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);
967 /* Beware: modes might be different */
968 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
969 if (mode == get_irn_mode(a)) {
972 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
974 } else if (is_Add(a)) {
975 if (mode_wrap_around(mode)) {
976 ir_node *left = get_Add_left(a);
977 ir_node *right = get_Add_right(a);
980 if (mode == get_irn_mode(right)) {
982 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
984 } else if (right == b) {
985 if (mode == get_irn_mode(left)) {
987 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
991 } else if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
992 ir_mode *mode = get_irn_mode(a);
994 if (mode == get_irn_mode(b)) {
1000 /* check if it's allowed to skip the conv */
1001 ma = get_irn_mode(a);
1002 mb = get_irn_mode(b);
1004 if (mode_is_reference(ma) && mode_is_reference(mb)) {
1005 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
1007 set_Sub_right(n, b);
1014 } /* equivalent_node_Sub */
1018 * Optimize an "idempotent unary op", ie op(op(n)) = n.
1021 * -(-a) == a, but might overflow two times.
1022 * We handle it anyway here but the better way would be a
1023 * flag. This would be needed for Pascal for instance.
1025 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
1027 ir_node *pred = get_unop_op(n);
1029 /* optimize symmetric unop */
1030 if (get_irn_op(pred) == get_irn_op(n)) {
1031 n = get_unop_op(pred);
1032 DBG_OPT_ALGSIM2(oldn, pred, n);
1035 } /* equivalent_node_idempotent_unop */
1037 /** Optimize Not(Not(x)) == x. */
1038 #define equivalent_node_Not equivalent_node_idempotent_unop
1040 /** --x == x ??? Is this possible or can --x raise an
1041 out of bounds exception if min =! max? */
1042 #define equivalent_node_Minus equivalent_node_idempotent_unop
1045 * Optimize a * 1 = 1 * a = a.
1047 static ir_node *equivalent_node_Mul(ir_node *n) {
1049 ir_node *a = get_Mul_left(n);
1050 ir_node *b = get_Mul_right(n);
1052 /* Mul is commutative and has again an other neutral element. */
1053 if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) {
1055 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1056 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) {
1058 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1061 } /* equivalent_node_Mul */
1064 * Optimize a / 1 = a.
1066 static ir_node *equivalent_node_Div(ir_node *n) {
1067 ir_node *a = get_Div_left(n);
1068 ir_node *b = get_Div_right(n);
1070 /* Div is not commutative. */
1071 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1072 /* Turn Div into a tuple (mem, bad, a) */
1073 ir_node *mem = get_Div_mem(n);
1074 ir_node *blk = get_irn_n(n, -1);
1075 turn_into_tuple(n, pn_Div_max);
1076 set_Tuple_pred(n, pn_Div_M, mem);
1077 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
1078 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
1079 set_Tuple_pred(n, pn_Div_res, a);
1082 } /* equivalent_node_Div */
1085 * Optimize a / 1.0 = a.
1087 static ir_node *equivalent_node_Quot(ir_node *n) {
1088 ir_node *a = get_Quot_left(n);
1089 ir_node *b = get_Quot_right(n);
1091 /* Div is not commutative. */
1092 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* Quot(x, 1) == x */
1093 /* Turn Quot into a tuple (mem, jmp, bad, a) */
1094 ir_node *mem = get_Quot_mem(n);
1095 ir_node *blk = get_irn_n(n, -1);
1096 turn_into_tuple(n, pn_Quot_max);
1097 set_Tuple_pred(n, pn_Quot_M, mem);
1098 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
1099 set_Tuple_pred(n, pn_Quot_X_except, new_Bad()); /* no exception */
1100 set_Tuple_pred(n, pn_Quot_res, a);
1103 } /* equivalent_node_Quot */
1106 * Optimize a / 1 = a.
1108 static ir_node *equivalent_node_DivMod(ir_node *n) {
1109 ir_node *b = get_DivMod_right(n);
1111 /* Div is not commutative. */
1112 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1113 /* Turn DivMod into a tuple (mem, jmp, bad, a, 0) */
1114 ir_node *a = get_DivMod_left(n);
1115 ir_node *mem = get_Div_mem(n);
1116 ir_node *blk = get_irn_n(n, -1);
1117 ir_mode *mode = get_DivMod_resmode(n);
1119 turn_into_tuple(n, pn_DivMod_max);
1120 set_Tuple_pred(n, pn_DivMod_M, mem);
1121 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
1122 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1123 set_Tuple_pred(n, pn_DivMod_res_div, a);
1124 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1127 } /* equivalent_node_DivMod */
1130 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1132 static ir_node *equivalent_node_Or(ir_node *n) {
1135 ir_node *a = get_Or_left(n);
1136 ir_node *b = get_Or_right(n);
1139 n = a; /* Or has it's own neutral element */
1140 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1141 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) {
1143 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1144 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
1146 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1150 } /* equivalent_node_Or */
1153 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a.
1155 static ir_node *equivalent_node_And(ir_node *n) {
1158 ir_node *a = get_And_left(n);
1159 ir_node *b = get_And_right(n);
1162 n = a; /* And has it's own neutral element */
1163 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1164 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) {
1166 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1167 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) {
1169 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1172 } /* equivalent_node_And */
1175 * Try to remove useless Conv's:
1177 static ir_node *equivalent_node_Conv(ir_node *n) {
1179 ir_node *a = get_Conv_op(n);
1182 ir_mode *n_mode = get_irn_mode(n);
1183 ir_mode *a_mode = get_irn_mode(a);
1185 if (n_mode == a_mode) { /* No Conv necessary */
1186 /* leave strict floating point Conv's */
1187 if (get_Conv_strict(n))
1190 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1191 } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */
1195 n_mode = get_irn_mode(n);
1196 b_mode = get_irn_mode(b);
1198 if (n_mode == b_mode) {
1199 if (n_mode == mode_b) {
1200 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1201 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1202 } else if (mode_is_int(n_mode) || mode_is_character(n_mode)) {
1203 if (smaller_mode(b_mode, a_mode)){
1204 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1205 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1211 } /* equivalent_node_Conv */
1214 * A Cast may be removed if the type of the previous node
1215 * is already the type of the Cast.
1217 static ir_node *equivalent_node_Cast(ir_node *n) {
1219 ir_node *pred = get_Cast_op(n);
1221 if (get_irn_type(pred) == get_Cast_type(n)) {
1223 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1226 } /* equivalent_node_Cast */
1229 * Several optimizations:
1230 * - no Phi in start block.
1231 * - remove Id operators that are inputs to Phi
1232 * - fold Phi-nodes, iff they have only one predecessor except
1235 static ir_node *equivalent_node_Phi(ir_node *n) {
1239 ir_node *block = NULL; /* to shutup gcc */
1240 ir_node *first_val = NULL; /* to shutup gcc */
1242 if (!get_opt_normalize()) return n;
1244 n_preds = get_Phi_n_preds(n);
1246 block = get_nodes_block(n);
1247 /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
1248 assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
1249 if ((is_Block_dead(block)) || /* Control dead */
1250 (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */
1251 return new_Bad(); /* in the Start Block. */
1253 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1255 /* If the Block has a Bad pred, we also have one. */
1256 for (i = 0; i < n_preds; ++i)
1257 if (is_Bad(get_Block_cfgpred(block, i)))
1258 set_Phi_pred(n, i, new_Bad());
1260 /* Find first non-self-referencing input */
1261 for (i = 0; i < n_preds; ++i) {
1262 first_val = get_Phi_pred(n, i);
1263 if ( (first_val != n) /* not self pointer */
1265 && (! is_Bad(first_val))
1267 ) { /* value not dead */
1268 break; /* then found first value. */
1273 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1277 /* search for rest of inputs, determine if any of these
1278 are non-self-referencing */
1279 while (++i < n_preds) {
1280 ir_node *scnd_val = get_Phi_pred(n, i);
1281 if ( (scnd_val != n)
1282 && (scnd_val != first_val)
1284 && (! is_Bad(scnd_val))
1292 /* Fold, if no multiple distinct non-self-referencing inputs */
1294 DBG_OPT_PHI(oldn, n);
1297 } /* equivalent_node_Phi */
1300 * Several optimizations:
1301 * - no Sync in start block.
1302 * - fold Sync-nodes, iff they have only one predecessor except
1305 static ir_node *equivalent_node_Sync(ir_node *n) {
1309 ir_node *first_val = NULL; /* to shutup gcc */
1311 if (!get_opt_normalize()) return n;
1313 n_preds = get_Sync_n_preds(n);
1315 /* Find first non-self-referencing input */
1316 for (i = 0; i < n_preds; ++i) {
1317 first_val = get_Sync_pred(n, i);
1318 if ((first_val != n) /* not self pointer */ &&
1319 (! is_Bad(first_val))
1320 ) { /* value not dead */
1321 break; /* then found first value. */
1326 /* A totally Bad or self-referencing Sync (we didn't break the above loop) */
1329 /* search the rest of inputs, determine if any of these
1330 are non-self-referencing */
1331 while (++i < n_preds) {
1332 ir_node *scnd_val = get_Sync_pred(n, i);
1333 if ((scnd_val != n) &&
1334 (scnd_val != first_val) &&
1335 (! is_Bad(scnd_val))
1341 /* Fold, if no multiple distinct non-self-referencing inputs */
1343 DBG_OPT_SYNC(oldn, n);
1346 } /* equivalent_node_Sync */
1349 * Optimize Proj(Tuple) and gigo() for ProjX in Bad block,
1350 * ProjX(Load) and ProjX(Store).
1352 static ir_node *equivalent_node_Proj(ir_node *proj) {
1353 ir_node *oldn = proj;
1354 ir_node *a = get_Proj_pred(proj);
1356 if (get_irn_op(a) == op_Tuple) {
1357 /* Remove the Tuple/Proj combination. */
1358 if ( get_Proj_proj(proj) <= get_Tuple_n_preds(a) ) {
1359 proj = get_Tuple_pred(a, get_Proj_proj(proj));
1360 DBG_OPT_TUPLE(oldn, a, proj);
1362 /* This should not happen! */
1363 assert(! "found a Proj with higher number than Tuple predecessors");
1366 } else if (get_irn_mode(proj) == mode_X) {
1367 if (is_Block_dead(get_nodes_block(skip_Proj(proj)))) {
1368 /* Remove dead control flow -- early gigo(). */
1370 } else if (get_opt_ldst_only_null_ptr_exceptions()) {
1371 ir_op *op = get_irn_op(a);
1373 if (op == op_Load) {
1374 /* get the Load address */
1375 ir_node *addr = get_Load_ptr(a);
1376 ir_node *blk = get_irn_n(a, -1);
1379 if (value_not_null(addr, &confirm)) {
1380 if (confirm == NULL) {
1381 /* this node may float if it did not depend on a Confirm */
1382 set_irn_pinned(a, op_pin_state_floats);
1384 if (get_Proj_proj(proj) == pn_Load_X_except) {
1385 DBG_OPT_EXC_REM(proj);
1388 return new_r_Jmp(current_ir_graph, blk);
1390 } else if (op == op_Store) {
1391 /* get the load/store address */
1392 ir_node *addr = get_Store_ptr(a);
1393 ir_node *blk = get_irn_n(a, -1);
1396 if (value_not_null(addr, &confirm)) {
1397 if (confirm == NULL) {
1398 /* this node may float if it did not depend on a Confirm */
1399 set_irn_pinned(a, op_pin_state_floats);
1401 if (get_Proj_proj(proj) == pn_Store_X_except) {
1402 DBG_OPT_EXC_REM(proj);
1405 return new_r_Jmp(current_ir_graph, blk);
1412 } /* equivalent_node_Proj */
1417 static ir_node *equivalent_node_Id(ir_node *n) {
1422 } while (get_irn_op(n) == op_Id);
1424 DBG_OPT_ID(oldn, n);
1426 } /* equivalent_node_Id */
1431 static ir_node *equivalent_node_Mux(ir_node *n)
1433 ir_node *oldn = n, *sel = get_Mux_sel(n);
1434 tarval *ts = value_of(sel);
1436 /* Mux(true, f, t) == t */
1437 if (ts == tarval_b_true) {
1438 n = get_Mux_true(n);
1439 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1441 /* Mux(false, f, t) == f */
1442 else if (ts == tarval_b_false) {
1443 n = get_Mux_false(n);
1444 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1446 /* Mux(v, x, x) == x */
1447 else if (get_Mux_false(n) == get_Mux_true(n)) {
1448 n = get_Mux_true(n);
1449 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1451 else if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(get_irn_mode(n))) {
1452 ir_node *cmp = get_Proj_pred(sel);
1453 long proj_nr = get_Proj_proj(sel);
1454 ir_node *b = get_Mux_false(n);
1455 ir_node *a = get_Mux_true(n);
1458 * Note: normalization puts the constant on the right site,
1459 * so we check only one case.
1461 * Note further that these optimization work even for floating point
1462 * with NaN's because -NaN == NaN.
1463 * However, if +0 and -0 is handled differently, we cannot use the first one.
1465 if (get_irn_op(cmp) == op_Cmp && get_Cmp_left(cmp) == a) {
1466 if (classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
1467 /* Mux(a CMP 0, X, a) */
1468 if (get_irn_op(b) == op_Minus && get_Minus_op(b) == a) {
1469 /* Mux(a CMP 0, -a, a) */
1470 if (proj_nr == pn_Cmp_Eq) {
1471 /* Mux(a == 0, -a, a) ==> -a */
1473 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1474 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1475 /* Mux(a != 0, -a, a) ==> a */
1477 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1479 } else if (classify_Const(b) == CNST_NULL) {
1480 /* Mux(a CMP 0, 0, a) */
1481 if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1482 /* Mux(a != 0, 0, a) ==> a */
1484 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1485 } else if (proj_nr == pn_Cmp_Eq) {
1486 /* Mux(a == 0, 0, a) ==> 0 */
1488 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1495 } /* equivalent_node_Mux */
1498 * Returns a equivalent node of a Psi: if a condition is true
1499 * and all previous conditions are false we know its value.
1500 * If all conditions are false its value is the default one.
1502 static ir_node *equivalent_node_Psi(ir_node *n) {
1504 return equivalent_node_Mux(n);
1506 } /* equivalent_node_Psi */
1509 * Optimize -a CMP -b into b CMP a.
1510 * This works only for for modes where unary Minus
1512 * Note that two-complement integers can Overflow
1513 * so it will NOT work.
1515 * For == and != can be handled in Proj(Cmp)
1517 static ir_node *equivalent_node_Cmp(ir_node *n) {
1518 ir_node *left = get_Cmp_left(n);
1519 ir_node *right = get_Cmp_right(n);
1521 if (get_irn_op(left) == op_Minus && get_irn_op(right) == op_Minus &&
1522 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1523 left = get_Minus_op(left);
1524 right = get_Minus_op(right);
1525 set_Cmp_left(n, right);
1526 set_Cmp_right(n, left);
1529 } /* equivalent_node_Cmp */
1532 * Remove Confirm nodes if setting is on.
1533 * Replace Confirms(x, '=', Constlike) by Constlike.
1535 static ir_node *equivalent_node_Confirm(ir_node *n) {
1536 ir_node *pred = get_Confirm_value(n);
1537 pn_Cmp pnc = get_Confirm_cmp(n);
1539 if (get_irn_op(pred) == op_Confirm && pnc == get_Confirm_cmp(pred)) {
1541 * rare case: two identical Confirms one after another,
1542 * replace the second one with the first.
1546 if (pnc == pn_Cmp_Eq) {
1547 ir_node *bound = get_Confirm_bound(n);
1550 * Optimize a rare case:
1551 * Confirm(x, '=', Constlike) ==> Constlike
1553 if (is_irn_constlike(bound)) {
1554 DBG_OPT_CONFIRM(n, bound);
1558 return get_opt_remove_confirm() ? get_Confirm_value(n) : n;
1562 * Optimize CopyB(mem, x, x) into a Nop.
1564 static ir_node *equivalent_node_CopyB(ir_node *n) {
1565 ir_node *a = get_CopyB_dst(n);
1566 ir_node *b = get_CopyB_src(n);
1569 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1570 ir_node *mem = get_CopyB_mem(n);
1571 ir_node *blk = get_nodes_block(n);
1572 turn_into_tuple(n, pn_CopyB_max);
1573 set_Tuple_pred(n, pn_CopyB_M, mem);
1574 set_Tuple_pred(n, pn_CopyB_X_regular, new_r_Jmp(current_ir_graph, blk));
1575 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1576 set_Tuple_pred(n, pn_CopyB_M_except, new_Bad());
1579 } /* equivalent_node_CopyB */
1582 * Optimize Bounds(idx, idx, upper) into idx.
1584 static ir_node *equivalent_node_Bound(ir_node *n) {
1585 ir_node *idx = get_Bound_index(n);
1586 ir_node *lower = get_Bound_lower(n);
1589 /* By definition lower < upper, so if idx == lower -->
1590 lower <= idx && idx < upper */
1592 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1595 ir_node *pred = skip_Proj(idx);
1597 if (get_irn_op(pred) == op_Bound) {
1599 * idx was Bounds_check previously, it is still valid if
1600 * lower <= pred_lower && pred_upper <= upper.
1602 ir_node *upper = get_Bound_upper(n);
1603 if (get_Bound_lower(pred) == lower &&
1604 get_Bound_upper(pred) == upper) {
1606 * One could expect that we simply return the previous
1607 * Bound here. However, this would be wrong, as we could
1608 * add an exception Proj to a new location than.
1609 * So, we must turn in into a tuple
1616 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1617 ir_node *mem = get_Bound_mem(n);
1618 ir_node *blk = get_nodes_block(n);
1619 turn_into_tuple(n, pn_Bound_max);
1620 set_Tuple_pred(n, pn_Bound_M, mem);
1621 set_Tuple_pred(n, pn_Bound_X_regular, new_r_Jmp(current_ir_graph, blk)); /* no exception */
1622 set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */
1623 set_Tuple_pred(n, pn_Bound_res, idx);
1626 } /* equivalent_node_Bound */
1629 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1630 * perform no actual computation, as, e.g., the Id nodes. It does not create
1631 * new nodes. It is therefore safe to free n if the node returned is not n.
1632 * If a node returns a Tuple we can not just skip it. If the size of the
1633 * in array fits, we transform n into a tuple (e.g., Div).
1635 ir_node *equivalent_node(ir_node *n) {
1636 if (n->op->ops.equivalent_node)
1637 return n->op->ops.equivalent_node(n);
1639 } /* equivalent_node */
1642 * Sets the default equivalent node operation for an ir_op_ops.
1644 * @param code the opcode for the default operation
1645 * @param ops the operations initialized
1650 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1654 ops->equivalent_node = equivalent_node_##a; \
1694 } /* firm_set_default_equivalent_node */
1697 * Do node specific optimizations of nodes predecessors.
1699 static void optimize_preds(ir_node *n) {
1700 switch (get_irn_opcode(n)) {
1702 case iro_Cmp: { /* We don't want Cast as input to Cmp. */
1703 ir_node *a = get_Cmp_left(n), *b = get_Cmp_right(n);
1705 if (get_irn_op(a) == op_Cast) {
1709 if (get_irn_op(b) == op_Cast) {
1711 set_Cmp_right(n, b);
1718 } /* optimize_preds */
1721 * Returns non-zero if a node is a Phi node
1722 * with all predecessors constant.
1724 static int is_const_Phi(ir_node *n) {
1729 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1730 if (! is_Const(get_irn_n(n, i)))
1733 } /* is_const_Phi */
1736 * Apply an evaluator on a binop with a constant operators (and one Phi).
1738 * @param phi the Phi node
1739 * @param other the other operand
1740 * @param eval an evaluator function
1741 * @param left if non-zero, other is the left operand, else the right
1743 * @return a new Phi node if the conversion was successful, NULL else
1745 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), int left) {
1751 int i, n = get_irn_arity(phi);
1753 NEW_ARR_A(void *, res, n);
1755 for (i = 0; i < n; ++i) {
1756 pred = get_irn_n(phi, i);
1757 tv = get_Const_tarval(pred);
1758 tv = eval(other, tv);
1760 if (tv == tarval_bad) {
1761 /* folding failed, bad */
1767 for (i = 0; i < n; ++i) {
1768 pred = get_irn_n(phi, i);
1769 tv = get_Const_tarval(pred);
1770 tv = eval(tv, other);
1772 if (tv == tarval_bad) {
1773 /* folding failed, bad */
1779 mode = get_irn_mode(phi);
1780 irg = current_ir_graph;
1781 for (i = 0; i < n; ++i) {
1782 pred = get_irn_n(phi, i);
1783 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1784 mode, res[i], get_Const_type(pred));
1786 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1787 } /* apply_binop_on_phi */
1790 * Apply an evaluator on a unop with a constant operator (a Phi).
1792 * @param phi the Phi node
1793 * @param eval an evaluator function
1795 * @return a new Phi node if the conversion was successful, NULL else
1797 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
1803 int i, n = get_irn_arity(phi);
1805 NEW_ARR_A(void *, res, n);
1806 for (i = 0; i < n; ++i) {
1807 pred = get_irn_n(phi, i);
1808 tv = get_Const_tarval(pred);
1811 if (tv == tarval_bad) {
1812 /* folding failed, bad */
1817 mode = get_irn_mode(phi);
1818 irg = current_ir_graph;
1819 for (i = 0; i < n; ++i) {
1820 pred = get_irn_n(phi, i);
1821 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1822 mode, res[i], get_Const_type(pred));
1824 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1825 } /* apply_unop_on_phi */
1828 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1829 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1830 * If possible, remove the Conv's.
1832 static ir_node *transform_node_AddSub(ir_node *n) {
1833 ir_mode *mode = get_irn_mode(n);
1835 if (mode_is_reference(mode)) {
1836 ir_node *left = get_binop_left(n);
1837 ir_node *right = get_binop_right(n);
1838 int ref_bits = get_mode_size_bits(mode);
1840 if (get_irn_op(left) == op_Conv) {
1841 ir_mode *mode = get_irn_mode(left);
1842 int bits = get_mode_size_bits(mode);
1844 if (ref_bits == bits &&
1845 mode_is_int(mode) &&
1846 get_mode_arithmetic(mode) == irma_twos_complement) {
1847 ir_node *pre = get_Conv_op(left);
1848 ir_mode *pre_mode = get_irn_mode(pre);
1850 if (mode_is_int(pre_mode) &&
1851 get_mode_size_bits(pre_mode) == bits &&
1852 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1853 /* ok, this conv just changes to sign, moreover the calculation
1854 * is done with same number of bits as our address mode, so
1855 * we can ignore the conv as address calculation can be viewed
1856 * as either signed or unsigned
1858 set_binop_left(n, pre);
1863 if (get_irn_op(right) == op_Conv) {
1864 ir_mode *mode = get_irn_mode(right);
1865 int bits = get_mode_size_bits(mode);
1867 if (ref_bits == bits &&
1868 mode_is_int(mode) &&
1869 get_mode_arithmetic(mode) == irma_twos_complement) {
1870 ir_node *pre = get_Conv_op(right);
1871 ir_mode *pre_mode = get_irn_mode(pre);
1873 if (mode_is_int(pre_mode) &&
1874 get_mode_size_bits(pre_mode) == bits &&
1875 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1876 /* ok, this conv just changes to sign, moreover the calculation
1877 * is done with same number of bits as our address mode, so
1878 * we can ignore the conv as address calculation can be viewed
1879 * as either signed or unsigned
1881 set_binop_right(n, pre);
1887 } /* transform_node_AddSub */
1889 #define HANDLE_BINOP_PHI(op,a,b,c) \
1891 if (is_Const(b) && is_const_Phi(a)) { \
1892 /* check for Op(Phi, Const) */ \
1893 c = apply_binop_on_phi(a, get_Const_tarval(b), op, 0); \
1895 else if (is_Const(a) && is_const_Phi(b)) { \
1896 /* check for Op(Const, Phi) */ \
1897 c = apply_binop_on_phi(b, get_Const_tarval(a), op, 1); \
1900 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1904 #define HANDLE_UNOP_PHI(op,a,c) \
1906 if (is_const_Phi(a)) { \
1907 /* check for Op(Phi) */ \
1908 c = apply_unop_on_phi(a, op); \
1911 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1917 * Do the AddSub optimization, then Transform
1918 * Constant folding on Phi
1919 * Add(a,a) -> Mul(a, 2)
1920 * Add(Mul(a, x), a) -> Mul(a, x+1)
1921 * if the mode is integer or float.
1922 * Transform Add(a,-b) into Sub(a,b).
1923 * Reassociation might fold this further.
1925 static ir_node *transform_node_Add(ir_node *n) {
1927 ir_node *a, *b, *c, *oldn = n;
1929 n = transform_node_AddSub(n);
1931 a = get_Add_left(n);
1932 b = get_Add_right(n);
1934 HANDLE_BINOP_PHI(tarval_add, a,b,c);
1936 mode = get_irn_mode(n);
1938 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1939 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1942 if (mode_is_num(mode)) {
1944 ir_node *block = get_irn_n(n, -1);
1947 get_irn_dbg_info(n),
1951 new_r_Const_long(current_ir_graph, block, mode, 2),
1953 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
1954 } else if (get_irn_op(a) == op_Minus) {
1956 get_irn_dbg_info(n),
1962 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1963 } else if (get_irn_op(b) == op_Minus) {
1965 get_irn_dbg_info(n),
1971 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1973 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1974 else if (!get_opt_reassociation() && get_irn_op(a) == op_Mul) {
1975 ir_node *ma = get_Mul_left(a);
1976 ir_node *mb = get_Mul_right(a);
1979 ir_node *blk = get_irn_n(n, -1);
1981 get_irn_dbg_info(n), current_ir_graph, blk,
1984 get_irn_dbg_info(n), current_ir_graph, blk,
1986 new_r_Const_long(current_ir_graph, blk, mode, 1),
1989 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1990 } else if (b == mb) {
1991 ir_node *blk = get_irn_n(n, -1);
1993 get_irn_dbg_info(n), current_ir_graph, blk,
1996 get_irn_dbg_info(n), current_ir_graph, blk,
1998 new_r_Const_long(current_ir_graph, blk, mode, 1),
2001 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
2004 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2005 else if (!get_opt_reassociation() && get_irn_op(b) == op_Mul) {
2006 ir_node *ma = get_Mul_left(b);
2007 ir_node *mb = get_Mul_right(b);
2010 ir_node *blk = get_irn_n(n, -1);
2012 get_irn_dbg_info(n), current_ir_graph, blk,
2015 get_irn_dbg_info(n), current_ir_graph, blk,
2017 new_r_Const_long(current_ir_graph, blk, mode, 1),
2020 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
2021 } else if (a == mb) {
2022 ir_node *blk = get_irn_n(n, -1);
2024 get_irn_dbg_info(n), current_ir_graph, blk,
2027 get_irn_dbg_info(n), current_ir_graph, blk,
2029 new_r_Const_long(current_ir_graph, blk, mode, 1),
2032 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
2037 } /* transform_node_Add */
2040 * Do the AddSub optimization, then Transform
2041 * Constant folding on Phi
2042 * Sub(0,a) -> Minus(a)
2043 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2044 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2046 static ir_node *transform_node_Sub(ir_node *n) {
2051 n = transform_node_AddSub(n);
2053 a = get_Sub_left(n);
2054 b = get_Sub_right(n);
2056 HANDLE_BINOP_PHI(tarval_sub, a,b,c);
2058 mode = get_irn_mode(n);
2060 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2061 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2064 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2065 if (mode_is_num(mode) && mode == get_irn_mode(a) && (classify_Const(a) == CNST_NULL)) {
2067 get_irn_dbg_info(n),
2072 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2074 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2075 else if (get_opt_reassociation() && get_irn_op(a) == op_Mul) {
2076 ir_node *ma = get_Mul_left(a);
2077 ir_node *mb = get_Mul_right(a);
2080 ir_node *blk = get_irn_n(n, -1);
2082 get_irn_dbg_info(n),
2083 current_ir_graph, blk,
2086 get_irn_dbg_info(n),
2087 current_ir_graph, blk,
2089 new_r_Const_long(current_ir_graph, blk, mode, 1),
2092 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2093 } else if (mb == b) {
2094 ir_node *blk = get_irn_n(n, -1);
2096 get_irn_dbg_info(n),
2097 current_ir_graph, blk,
2100 get_irn_dbg_info(n),
2101 current_ir_graph, blk,
2103 new_r_Const_long(current_ir_graph, blk, mode, 1),
2106 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2108 } else if (get_irn_op(a) == op_Sub) {
2109 ir_node *x = get_Sub_left(a);
2110 ir_node *y = get_Sub_right(a);
2111 ir_node *blk = get_irn_n(n, -1);
2112 ir_mode *m_b = get_irn_mode(b);
2113 ir_mode *m_y = get_irn_mode(y);
2116 /* Determine the right mode for the Add. */
2119 else if (mode_is_reference(m_b))
2121 else if (mode_is_reference(m_y))
2125 * Both modes are different but none is reference,
2126 * happens for instance in SubP(SubP(P, Iu), Is).
2127 * We have two possibilities here: Cast or ignore.
2128 * Currently we ignore this case.
2133 add = new_r_Add(current_ir_graph, blk, y, b, mode);
2136 set_Sub_right(n, add);
2137 DBG_OPT_ALGSIM0(n, n, FS_OPT_SUB_SUB_X_Y_Z);
2141 } /* transform_node_Sub */
2144 * Transform Mul(a,-1) into -a.
2145 * Do constant evaluation of Phi nodes.
2146 * Do architecture dependent optimizations on Mul nodes
2148 static ir_node *transform_node_Mul(ir_node *n) {
2149 ir_node *c, *oldn = n;
2150 ir_node *a = get_Mul_left(n);
2151 ir_node *b = get_Mul_right(n);
2154 HANDLE_BINOP_PHI(tarval_mul, a,b,c);
2156 mode = get_irn_mode(n);
2157 if (mode_is_signed(mode)) {
2160 if (value_of(a) == get_mode_minus_one(mode))
2162 else if (value_of(b) == get_mode_minus_one(mode))
2165 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2166 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2170 return arch_dep_replace_mul_with_shifts(n);
2171 } /* transform_node_Mul */
2174 * Transform a Div Node.
2176 static ir_node *transform_node_Div(ir_node *n) {
2177 tarval *tv = value_of(n);
2180 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2182 if (tv != tarval_bad) {
2183 value = new_Const(get_tarval_mode(tv), tv);
2185 DBG_OPT_CSTEVAL(n, value);
2186 } else /* Try architecture dependent optimization */
2187 value = arch_dep_replace_div_by_const(n);
2190 /* Turn Div into a tuple (mem, jmp, bad, value) */
2191 ir_node *mem = get_Div_mem(n);
2192 ir_node *blk = get_irn_n(n, -1);
2194 turn_into_tuple(n, pn_Div_max);
2195 set_Tuple_pred(n, pn_Div_M, mem);
2196 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2197 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2198 set_Tuple_pred(n, pn_Div_res, value);
2201 } /* transform_node_Div */
2204 * Transform a Mod node.
2206 static ir_node *transform_node_Mod(ir_node *n) {
2207 tarval *tv = value_of(n);
2210 /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */
2212 if (tv != tarval_bad) {
2213 value = new_Const(get_tarval_mode(tv), tv);
2215 DBG_OPT_CSTEVAL(n, value);
2216 } else /* Try architecture dependent optimization */
2217 value = arch_dep_replace_mod_by_const(n);
2220 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2221 ir_node *mem = get_Mod_mem(n);
2222 ir_node *blk = get_irn_n(n, -1);
2224 turn_into_tuple(n, pn_Mod_max);
2225 set_Tuple_pred(n, pn_Mod_M, mem);
2226 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2227 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2228 set_Tuple_pred(n, pn_Mod_res, value);
2231 } /* transform_node_Mod */
2234 * Transform a DivMod node.
2236 static ir_node *transform_node_DivMod(ir_node *n) {
2239 ir_node *a = get_DivMod_left(n);
2240 ir_node *b = get_DivMod_right(n);
2241 ir_mode *mode = get_irn_mode(a);
2242 tarval *ta = value_of(a);
2243 tarval *tb = value_of(b);
2245 if (!(mode_is_int(mode) && mode_is_int(get_irn_mode(b))))
2248 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2250 if (tb != tarval_bad) {
2251 if (tb == get_mode_one(get_tarval_mode(tb))) {
2252 b = new_Const (mode, get_mode_null(mode));
2255 DBG_OPT_CSTEVAL(n, b);
2256 } else if (ta != tarval_bad) {
2257 tarval *resa, *resb;
2258 resa = tarval_div (ta, tb);
2259 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2260 Jmp for X result!? */
2261 resb = tarval_mod (ta, tb);
2262 if (resb == tarval_bad) return n; /* Causes exception! */
2263 a = new_Const (mode, resa);
2264 b = new_Const (mode, resb);
2267 DBG_OPT_CSTEVAL(n, a);
2268 DBG_OPT_CSTEVAL(n, b);
2269 } else { /* Try architecture dependent optimization */
2270 arch_dep_replace_divmod_by_const(&a, &b, n);
2271 evaluated = a != NULL;
2273 } else if (ta == get_mode_null(mode)) {
2274 /* 0 / non-Const = 0 */
2279 if (evaluated) { /* replace by tuple */
2280 ir_node *mem = get_DivMod_mem(n);
2281 ir_node *blk = get_irn_n(n, -1);
2282 turn_into_tuple(n, pn_DivMod_max);
2283 set_Tuple_pred(n, pn_DivMod_M, mem);
2284 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
2285 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
2286 set_Tuple_pred(n, pn_DivMod_res_div, a);
2287 set_Tuple_pred(n, pn_DivMod_res_mod, b);
2291 } /* transform_node_DivMod */
2294 * Optimize Abs(x) into x if x is Confirmed >= 0
2295 * Optimize Abs(x) into -x if x is Confirmed <= 0
2297 static ir_node *transform_node_Abs(ir_node *n) {
2299 ir_node *a = get_Abs_op(n);
2300 value_classify_sign sign = classify_value_sign(a);
2302 if (sign == value_classified_negative) {
2303 ir_mode *mode = get_irn_mode(n);
2306 * We can replace the Abs by -x here.
2307 * We even could add a new Confirm here.
2309 * Note that -x would create a new node, so we could
2310 * not run it in the equivalent_node() context.
2312 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
2313 get_irn_n(n, -1), a, mode);
2315 DBG_OPT_CONFIRM(oldn, n);
2316 } else if (sign == value_classified_positive) {
2317 /* n is positive, Abs is not needed */
2320 DBG_OPT_CONFIRM(oldn, n);
2324 } /* transform_node_Abs */
2327 * Transform a Cond node.
2329 * Replace the Cond by a Jmp if it branches on a constant
2332 static ir_node *transform_node_Cond(ir_node *n) {
2335 ir_node *a = get_Cond_selector(n);
2336 tarval *ta = value_of(a);
2338 /* we need block info which is not available in floating irgs */
2339 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
2342 if ((ta != tarval_bad) &&
2343 (get_irn_mode(a) == mode_b) &&
2344 (get_opt_unreachable_code())) {
2345 /* It's a boolean Cond, branching on a boolean constant.
2346 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2347 jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n));
2348 turn_into_tuple(n, pn_Cond_max);
2349 if (ta == tarval_b_true) {
2350 set_Tuple_pred(n, pn_Cond_false, new_Bad());
2351 set_Tuple_pred(n, pn_Cond_true, jmp);
2353 set_Tuple_pred(n, pn_Cond_false, jmp);
2354 set_Tuple_pred(n, pn_Cond_true, new_Bad());
2356 /* We might generate an endless loop, so keep it alive. */
2357 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
2360 } /* transform_node_Cond */
2365 static ir_node *transform_node_And(ir_node *n) {
2366 ir_node *c, *oldn = n;
2367 ir_node *a = get_And_left(n);
2368 ir_node *b = get_And_right(n);
2370 HANDLE_BINOP_PHI(tarval_and, a,b,c);
2372 } /* transform_node_And */
2377 static ir_node *transform_node_Eor(ir_node *n) {
2378 ir_node *c, *oldn = n;
2379 ir_node *a = get_Eor_left(n);
2380 ir_node *b = get_Eor_right(n);
2381 ir_mode *mode = get_irn_mode(n);
2383 HANDLE_BINOP_PHI(tarval_eor, a,b,c);
2387 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
2388 mode, get_mode_null(mode));
2389 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
2390 } else if ((mode == mode_b)
2391 && (get_irn_op(a) == op_Proj)
2392 && (get_irn_mode(a) == mode_b)
2393 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)
2394 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2395 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
2396 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2397 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
2399 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
2400 } else if ((mode == mode_b)
2401 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) {
2402 /* The Eor is a Not. Replace it by a Not. */
2403 /* ????!!!Extend to bitfield 1111111. */
2404 n = new_r_Not(current_ir_graph, get_irn_n(n, -1), a, mode_b);
2406 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2410 } /* transform_node_Eor */
2415 static ir_node *transform_node_Not(ir_node *n) {
2416 ir_node *c, *oldn = n;
2417 ir_node *a = get_Not_op(n);
2419 HANDLE_UNOP_PHI(tarval_not,a,c);
2421 /* check for a boolean Not */
2422 if ( (get_irn_mode(n) == mode_b)
2423 && (get_irn_op(a) == op_Proj)
2424 && (get_irn_mode(a) == mode_b)
2425 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2426 /* We negate a Cmp. The Cmp has the negated result anyways! */
2427 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2428 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
2429 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
2432 } /* transform_node_Not */
2435 * Transform a Minus.
2437 static ir_node *transform_node_Minus(ir_node *n) {
2438 ir_node *c, *oldn = n;
2439 ir_node *a = get_Minus_op(n);
2441 HANDLE_UNOP_PHI(tarval_neg,a,c);
2443 } /* transform_node_Minus */
2446 * Transform a Cast_type(Const) into a new Const_type
2448 static ir_node *transform_node_Cast(ir_node *n) {
2450 ir_node *pred = get_Cast_op(n);
2451 ir_type *tp = get_irn_type(n);
2453 if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) {
2454 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
2455 get_Const_tarval(pred), tp);
2456 DBG_OPT_CSTEVAL(oldn, n);
2457 } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) {
2458 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred),
2459 get_SymConst_kind(pred), tp);
2460 DBG_OPT_CSTEVAL(oldn, n);
2464 } /* transform_node_Cast */
2467 * Transform a Proj(Div) with a non-zero value.
2468 * Removes the exceptions and routes the memory to the NoMem node.
2470 static ir_node *transform_node_Proj_Div(ir_node *proj) {
2471 ir_node *div = get_Proj_pred(proj);
2472 ir_node *b = get_Div_right(div);
2473 ir_node *confirm, *res, *new_mem;
2476 if (value_not_zero(b, &confirm)) {
2477 /* div(x, y) && y != 0 */
2478 proj_nr = get_Proj_proj(proj);
2480 case pn_Div_X_regular:
2481 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
2483 case pn_Div_X_except:
2484 /* we found an exception handler, remove it */
2485 DBG_OPT_EXC_REM(proj);
2489 res = get_Div_mem(div);
2490 new_mem = get_irg_no_mem(current_ir_graph);
2493 /* This node can only float up to the Confirm block */
2494 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2496 set_irn_pinned(div, op_pin_state_floats);
2497 /* this is a Div without exception, we can remove the memory edge */
2498 set_Div_mem(div, new_mem);
2503 } /* transform_node_Proj_Div */
2506 * Transform a Proj(Mod) with a non-zero value.
2507 * Removes the exceptions and routes the memory to the NoMem node.
2509 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
2510 ir_node *mod = get_Proj_pred(proj);
2511 ir_node *b = get_Mod_right(mod);
2512 ir_node *confirm, *res, *new_mem;
2515 if (value_not_zero(b, &confirm)) {
2516 /* mod(x, y) && y != 0 */
2517 proj_nr = get_Proj_proj(proj);
2521 case pn_Mod_X_regular:
2522 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
2524 case pn_Mod_X_except:
2525 /* we found an exception handler, remove it */
2526 DBG_OPT_EXC_REM(proj);
2530 res = get_Mod_mem(mod);
2531 new_mem = get_irg_no_mem(current_ir_graph);
2534 /* This node can only float up to the Confirm block */
2535 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2537 set_irn_pinned(mod, op_pin_state_floats);
2538 /* this is a Mod without exception, we can remove the memory edge */
2539 set_Mod_mem(mod, get_irg_no_mem(current_ir_graph));
2542 if (get_Mod_left(mod) == b) {
2543 /* a % a = 0 if a != 0 */
2544 ir_mode *mode = get_irn_mode(proj);
2545 ir_node *res = new_Const(mode, get_mode_null(mode));
2547 DBG_OPT_CSTEVAL(mod, res);
2553 } /* transform_node_Proj_Mod */
2556 * Transform a Proj(DivMod) with a non-zero value.
2557 * Removes the exceptions and routes the memory to the NoMem node.
2559 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
2560 ir_node *divmod = get_Proj_pred(proj);
2561 ir_node *b = get_DivMod_right(divmod);
2562 ir_node *confirm, *res, *new_mem;
2565 if (value_not_zero(b, &confirm)) {
2566 /* DivMod(x, y) && y != 0 */
2567 proj_nr = get_Proj_proj(proj);
2571 case pn_DivMod_X_regular:
2572 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
2574 case pn_DivMod_X_except:
2575 /* we found an exception handler, remove it */
2576 DBG_OPT_EXC_REM(proj);
2580 res = get_DivMod_mem(divmod);
2581 new_mem = get_irg_no_mem(current_ir_graph);
2584 /* This node can only float up to the Confirm block */
2585 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2587 set_irn_pinned(divmod, op_pin_state_floats);
2588 /* this is a DivMod without exception, we can remove the memory edge */
2589 set_DivMod_mem(divmod, get_irg_no_mem(current_ir_graph));
2592 case pn_DivMod_res_mod:
2593 if (get_DivMod_left(divmod) == b) {
2594 /* a % a = 0 if a != 0 */
2595 ir_mode *mode = get_irn_mode(proj);
2596 ir_node *res = new_Const(mode, get_mode_null(mode));
2598 DBG_OPT_CSTEVAL(divmod, res);
2604 } /* transform_node_Proj_DivMod */
2607 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
2609 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
2610 if (get_opt_unreachable_code()) {
2611 ir_node *n = get_Proj_pred(proj);
2612 ir_node *b = get_Cond_selector(n);
2614 if (mode_is_int(get_irn_mode(b))) {
2615 tarval *tb = value_of(b);
2617 if (tb != tarval_bad) {
2618 /* we have a constant switch */
2619 long num = get_Proj_proj(proj);
2621 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
2622 if (get_tarval_long(tb) == num) {
2623 /* Do NOT create a jump here, or we will have 2 control flow ops
2624 * in a block. This case is optimized away in optimize_cf(). */
2627 /* this case will NEVER be taken, kill it */
2635 } /* transform_node_Proj_Cond */
2638 * Normalizes and optimizes Cmp nodes.
2640 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
2641 if (get_opt_reassociation()) {
2642 ir_node *n = get_Proj_pred(proj);
2643 ir_node *left = get_Cmp_left(n);
2644 ir_node *right = get_Cmp_right(n);
2648 ir_mode *mode = NULL;
2649 long proj_nr = get_Proj_proj(proj);
2652 * First step: normalize the compare op
2653 * by placing the constant on the right site
2654 * or moving the lower address node to the left.
2655 * We ignore the case that both are constants
2656 * this case should be optimized away.
2658 if (get_irn_op(right) == op_Const) {
2660 } else if (get_irn_op(left) == op_Const) {
2665 proj_nr = get_inversed_pnc(proj_nr);
2667 } else if (get_irn_idx(left) > get_irn_idx(right)) {
2673 proj_nr = get_inversed_pnc(proj_nr);
2678 * Second step: Try to reduce the magnitude
2679 * of a constant. This may help to generate better code
2680 * later and may help to normalize more compares.
2681 * Of course this is only possible for integer values.
2684 mode = get_irn_mode(c);
2685 tv = get_Const_tarval(c);
2687 if (tv != tarval_bad) {
2688 /* the following optimization is possible on modes without Overflow
2689 * on Unary Minus or on == and !=:
2690 * -a CMP c ==> a swap(CMP) -c
2692 * Beware: for two-complement Overflow may occur, so only == and != can
2693 * be optimized, see this:
2694 * -MININT < 0 =/=> MININT > 0 !!!
2696 if (get_opt_constant_folding() && get_irn_op(left) == op_Minus &&
2697 (!mode_overflow_on_unary_Minus(mode) ||
2698 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
2699 left = get_Minus_op(left);
2700 tv = tarval_sub(get_mode_null(mode), tv);
2702 proj_nr = get_inversed_pnc(proj_nr);
2706 /* for integer modes, we have more */
2707 if (mode_is_int(mode)) {
2708 /* Ne includes Unordered which is not possible on integers.
2709 * However, frontends often use this wrong, so fix it here */
2710 if (proj_nr & pn_Cmp_Uo) {
2711 proj_nr &= ~pn_Cmp_Uo;
2712 set_Proj_proj(proj, proj_nr);
2715 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
2716 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
2717 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
2718 tv = tarval_sub(tv, get_mode_one(mode));
2720 proj_nr ^= pn_Cmp_Eq;
2723 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
2724 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
2725 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
2726 tv = tarval_add(tv, get_mode_one(mode));
2728 proj_nr ^= pn_Cmp_Eq;
2732 /* the following reassociations work only for == and != */
2733 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
2735 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
2736 if (classify_tarval(tv) == TV_CLASSIFY_NULL && get_irn_op(left) == op_Sub) {
2737 right = get_Sub_right(left);
2738 left = get_Sub_left(left);
2740 tv = value_of(right);
2744 if (tv != tarval_bad) {
2745 ir_op *op = get_irn_op(left);
2747 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
2749 ir_node *c1 = get_Sub_right(left);
2750 tarval *tv2 = value_of(c1);
2752 if (tv2 != tarval_bad) {
2753 tv2 = tarval_add(tv, value_of(c1));
2755 if (tv2 != tarval_bad) {
2756 left = get_Sub_left(left);
2762 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
2763 else if (op == op_Add) {
2764 ir_node *a_l = get_Add_left(left);
2765 ir_node *a_r = get_Add_right(left);
2769 if (get_irn_op(a_l) == op_Const) {
2771 tv2 = value_of(a_l);
2774 tv2 = value_of(a_r);
2777 if (tv2 != tarval_bad) {
2778 tv2 = tarval_sub(tv, tv2);
2780 if (tv2 != tarval_bad) {
2787 /* -a == c ==> a == -c, -a != c ==> a != -c */
2788 else if (op == op_Minus) {
2789 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
2791 if (tv2 != tarval_bad) {
2792 left = get_Minus_op(left);
2799 /* the following reassociations work only for <= */
2800 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2801 if (tv != tarval_bad) {
2802 ir_op *op = get_irn_op(left);
2804 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
2812 * optimization for AND:
2814 * And(x, C) == C ==> And(x, C) != 0
2815 * And(x, C) != C ==> And(X, C) == 0
2817 * if C is a single Bit constant.
2819 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) &&
2820 (get_irn_op(left) == op_And)) {
2821 if (is_single_bit_tarval(tv)) {
2822 /* check for Constant's match. We have check hare the tarvals,
2823 because our const might be changed */
2824 ir_node *la = get_And_left(left);
2825 ir_node *ra = get_And_right(left);
2826 if ((is_Const(la) && get_Const_tarval(la) == tv) ||
2827 (is_Const(ra) && get_Const_tarval(ra) == tv)) {
2828 /* fine: do the transformation */
2829 tv = get_mode_null(get_tarval_mode(tv));
2830 proj_nr ^= pn_Cmp_Leg;
2835 } /* tarval != bad */
2839 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
2841 if (changed & 2) /* need a new Const */
2842 right = new_Const(mode, tv);
2844 /* create a new compare */
2845 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block,
2848 set_Proj_pred(proj, n);
2849 set_Proj_proj(proj, proj_nr);
2853 } /* transform_node_Proj_Cmp */
2856 * Does all optimizations on nodes that must be done on it's Proj's
2857 * because of creating new nodes.
2859 static ir_node *transform_node_Proj(ir_node *proj) {
2860 ir_node *n = get_Proj_pred(proj);
2862 switch (get_irn_opcode(n)) {
2864 return transform_node_Proj_Div(proj);
2867 return transform_node_Proj_Mod(proj);
2870 return transform_node_Proj_DivMod(proj);
2873 return transform_node_Proj_Cond(proj);
2876 return transform_node_Proj_Cmp(proj);
2879 /* should not happen, but if it does will be optimized away */
2880 return equivalent_node_Proj(proj);
2886 } /* transform_node_Proj */
2889 * Move Confirms down through Phi nodes.
2891 static ir_node *transform_node_Phi(ir_node *phi) {
2893 ir_mode *mode = get_irn_mode(phi);
2895 if (mode_is_reference(mode)) {
2896 n = get_irn_arity(phi);
2898 /* Beware of Phi0 */
2900 ir_node *pred = get_irn_n(phi, 0);
2901 ir_node *bound, *new_Phi, *block, **in;
2904 if (! is_Confirm(pred))
2907 bound = get_Confirm_bound(pred);
2908 pnc = get_Confirm_cmp(pred);
2910 NEW_ARR_A(ir_node *, in, n);
2911 in[0] = get_Confirm_value(pred);
2913 for (i = 1; i < n; ++i) {
2914 pred = get_irn_n(phi, i);
2916 if (! is_Confirm(pred) ||
2917 get_Confirm_bound(pred) != bound ||
2918 get_Confirm_cmp(pred) != pnc)
2920 in[i] = get_Confirm_value(pred);
2922 /* move the Confirm nodes "behind" the Phi */
2923 block = get_irn_n(phi, -1);
2924 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
2925 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
2929 } /* transform_node_Phi */
2932 * Returns the operands of a commutative bin-op, if one operand is
2933 * a const, it is returned as the second one.
2935 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
2936 ir_node *op_a = get_binop_left(binop);
2937 ir_node *op_b = get_binop_right(binop);
2939 assert(is_op_commutative(get_irn_op(binop)));
2941 if (get_irn_op(op_a) == op_Const) {
2948 } /* get_comm_Binop_Ops */
2951 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
2952 * Such pattern may arise in bitfield stores.
2954 * value c4 value c4 & c2
2955 * AND c3 AND c1 | c3
2960 static ir_node *transform_node_Or_bf_store(ir_node *or) {
2963 ir_node *and_l, *c3;
2964 ir_node *value, *c4;
2965 ir_node *new_and, *new_const, *block;
2966 ir_mode *mode = get_irn_mode(or);
2968 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
2970 get_comm_Binop_Ops(or, &and, &c1);
2971 if ((get_irn_op(c1) != op_Const) || (get_irn_op(and) != op_And))
2974 get_comm_Binop_Ops(and, &or_l, &c2);
2975 if ((get_irn_op(c2) != op_Const) || (get_irn_op(or_l) != op_Or))
2978 get_comm_Binop_Ops(or_l, &and_l, &c3);
2979 if ((get_irn_op(c3) != op_Const) || (get_irn_op(and_l) != op_And))
2982 get_comm_Binop_Ops(and_l, &value, &c4);
2983 if (get_irn_op(c4) != op_Const)
2986 /* ok, found the pattern, check for conditions */
2987 assert(mode == get_irn_mode(and));
2988 assert(mode == get_irn_mode(or_l));
2989 assert(mode == get_irn_mode(and_l));
2991 tv1 = get_Const_tarval(c1);
2992 tv2 = get_Const_tarval(c2);
2993 tv3 = get_Const_tarval(c3);
2994 tv4 = get_Const_tarval(c4);
2996 tv = tarval_or(tv4, tv2);
2997 if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) {
2998 /* have at least one 0 at the same bit position */
3002 n_tv4 = tarval_not(tv4);
3003 if (tv3 != tarval_and(tv3, n_tv4)) {
3004 /* bit in the or_mask is outside the and_mask */
3008 n_tv2 = tarval_not(tv2);
3009 if (tv1 != tarval_and(tv1, n_tv2)) {
3010 /* bit in the or_mask is outside the and_mask */
3014 /* ok, all conditions met */
3015 block = get_irn_n(or, -1);
3017 new_and = new_r_And(current_ir_graph, block,
3018 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
3020 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
3022 set_Or_left(or, new_and);
3023 set_Or_right(or, new_const);
3025 /* check for more */
3026 return transform_node_Or_bf_store(or);
3027 } /* transform_node_Or_bf_store */
3030 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
3032 static ir_node *transform_node_Or_Rot(ir_node *or) {
3033 ir_mode *mode = get_irn_mode(or);
3034 ir_node *shl, *shr, *block;
3035 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
3038 if (! mode_is_int(mode))
3041 shl = get_binop_left(or);
3042 shr = get_binop_right(or);
3044 if (get_irn_op(shl) == op_Shr) {
3045 if (get_irn_op(shr) != op_Shl)
3051 } else if (get_irn_op(shl) != op_Shl) {
3053 } else if (get_irn_op(shr) != op_Shr) {
3056 x = get_Shl_left(shl);
3057 if (x != get_Shr_left(shr))
3060 c1 = get_Shl_right(shl);
3061 c2 = get_Shr_right(shr);
3062 if (get_irn_op(c1) == op_Const && get_irn_op(c2) == op_Const) {
3063 tv1 = get_Const_tarval(c1);
3064 if (! tarval_is_long(tv1))
3067 tv2 = get_Const_tarval(c2);
3068 if (! tarval_is_long(tv2))
3071 if (get_tarval_long(tv1) + get_tarval_long(tv2)
3072 != get_mode_size_bits(mode))
3075 /* yet, condition met */
3076 block = get_irn_n(or, -1);
3078 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
3080 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
3082 } else if (get_irn_op(c1) == op_Sub) {
3086 if (get_Sub_right(sub) != v)
3089 c1 = get_Sub_left(sub);
3090 if (get_irn_op(c1) != op_Const)
3093 tv1 = get_Const_tarval(c1);
3094 if (! tarval_is_long(tv1))
3097 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
3100 /* yet, condition met */
3101 block = get_nodes_block(or);
3103 /* a Rot right is not supported, so use a rot left */
3104 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
3106 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
3108 } else if (get_irn_op(c2) == op_Sub) {
3112 c1 = get_Sub_left(sub);
3113 if (get_irn_op(c1) != op_Const)
3116 tv1 = get_Const_tarval(c1);
3117 if (! tarval_is_long(tv1))
3120 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
3123 /* yet, condition met */
3124 block = get_irn_n(or, -1);
3127 n = new_r_Rot(current_ir_graph, block, x, v, mode);
3129 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
3134 } /* transform_node_Or_Rot */
3139 static ir_node *transform_node_Or(ir_node *n) {
3140 ir_node *c, *oldn = n;
3141 ir_node *a = get_Or_left(n);
3142 ir_node *b = get_Or_right(n);
3144 HANDLE_BINOP_PHI(tarval_or, a,b,c);
3146 n = transform_node_Or_bf_store(n);
3147 n = transform_node_Or_Rot(n);
3150 } /* transform_node_Or */
3154 static ir_node *transform_node(ir_node *n);
3157 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl.
3159 * Should be moved to reassociation?
3161 static ir_node *transform_node_shift(ir_node *n) {
3162 ir_node *left, *right;
3163 tarval *tv1, *tv2, *res;
3165 int modulo_shf, flag;
3167 left = get_binop_left(n);
3169 /* different operations */
3170 if (get_irn_op(left) != get_irn_op(n))
3173 right = get_binop_right(n);
3174 tv1 = value_of(right);
3175 if (tv1 == tarval_bad)
3178 tv2 = value_of(get_binop_right(left));
3179 if (tv2 == tarval_bad)
3182 res = tarval_add(tv1, tv2);
3184 /* beware: a simple replacement works only, if res < modulo shift */
3185 mode = get_irn_mode(n);
3189 modulo_shf = get_mode_modulo_shift(mode);
3190 if (modulo_shf > 0) {
3191 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
3193 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
3199 /* ok, we can replace it */
3200 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
3202 in[0] = get_binop_left(left);
3203 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
3205 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
3207 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
3209 return transform_node(irn);
3212 } /* transform_node_shift */
3217 static ir_node *transform_node_Shr(ir_node *n) {
3218 ir_node *c, *oldn = n;
3219 ir_node *a = get_Shr_left(n);
3220 ir_node *b = get_Shr_right(n);
3222 HANDLE_BINOP_PHI(tarval_shr, a, b, c);
3223 return transform_node_shift(n);
3224 } /* transform_node_Shr */
3229 static ir_node *transform_node_Shrs(ir_node *n) {
3230 ir_node *c, *oldn = n;
3231 ir_node *a = get_Shrs_left(n);
3232 ir_node *b = get_Shrs_right(n);
3234 HANDLE_BINOP_PHI(tarval_shrs, a, b, c);
3235 return transform_node_shift(n);
3236 } /* transform_node_Shrs */
3241 static ir_node *transform_node_Shl(ir_node *n) {
3242 ir_node *c, *oldn = n;
3243 ir_node *a = get_Shl_left(n);
3244 ir_node *b = get_Shl_right(n);
3246 HANDLE_BINOP_PHI(tarval_shl, a, b, c);
3247 return transform_node_shift(n);
3248 } /* transform_node_Shl */
3251 * Remove dead blocks and nodes in dead blocks
3252 * in keep alive list. We do not generate a new End node.
3254 static ir_node *transform_node_End(ir_node *n) {
3255 int i, j, n_keepalives = get_End_n_keepalives(n);
3258 NEW_ARR_A(ir_node *, in, n_keepalives);
3260 for (i = j = 0; i < n_keepalives; ++i) {
3261 ir_node *ka = get_End_keepalive(n, i);
3263 if (! is_Block_dead(ka)) {
3267 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
3270 /* FIXME: beabi need to keep a Proj(M) */
3271 if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka))
3274 if (j != n_keepalives)
3275 set_End_keepalives(n, j, in);
3277 } /* transform_node_End */
3280 * Optimize a Mux into some simpler cases.
3282 static ir_node *transform_node_Mux(ir_node *n) {
3283 ir_node *oldn = n, *sel = get_Mux_sel(n);
3284 ir_mode *mode = get_irn_mode(n);
3286 if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(mode)) {
3287 ir_node *cmp = get_Proj_pred(sel);
3288 long proj_nr = get_Proj_proj(sel);
3289 ir_node *f = get_Mux_false(n);
3290 ir_node *t = get_Mux_true(n);
3292 if (get_irn_op(cmp) == op_Cmp && classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
3293 ir_node *block = get_irn_n(n, -1);
3296 * Note: normalization puts the constant on the right site,
3297 * so we check only one case.
3299 * Note further that these optimization work even for floating point
3300 * with NaN's because -NaN == NaN.
3301 * However, if +0 and -0 is handled differently, we cannot use the first one.
3303 if (get_irn_op(f) == op_Minus &&
3304 get_Minus_op(f) == t &&
3305 get_Cmp_left(cmp) == t) {
3307 if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3308 /* Mux(a >=/> 0, -a, a) ==> Abs(a) */
3309 n = new_rd_Abs(get_irn_dbg_info(n),
3313 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3315 } else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3316 /* Mux(a <=/< 0, -a, a) ==> Minus(Abs(a)) */
3317 n = new_rd_Abs(get_irn_dbg_info(n),
3321 n = new_rd_Minus(get_irn_dbg_info(n),
3326 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3329 } else if (get_irn_op(t) == op_Minus &&
3330 get_Minus_op(t) == f &&
3331 get_Cmp_left(cmp) == f) {
3333 if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3334 /* Mux(a <=/< 0, a, -a) ==> Abs(a) */
3335 n = new_rd_Abs(get_irn_dbg_info(n),
3339 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3341 } else if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3342 /* Mux(a >=/> 0, a, -a) ==> Minus(Abs(a)) */
3343 n = new_rd_Abs(get_irn_dbg_info(n),
3347 n = new_rd_Minus(get_irn_dbg_info(n),
3352 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3357 if (mode_is_int(mode) && mode_is_signed(mode) &&
3358 get_mode_arithmetic(mode) == irma_twos_complement) {
3359 ir_node *x = get_Cmp_left(cmp);
3361 /* the following optimization works only with signed integer two-complement mode */
3363 if (mode == get_irn_mode(x)) {
3365 * FIXME: this restriction is two rigid, as it would still
3366 * work if mode(x) = Hs and mode == Is, but at least it removes
3369 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Le) &&
3370 classify_Const(t) == CNST_ALL_ONE &&
3371 classify_Const(f) == CNST_NULL) {
3373 * Mux(x:T </<= 0, 0, -1) -> Shrs(x, sizeof_bits(T) - 1)
3377 n = new_rd_Shrs(get_irn_dbg_info(n),
3378 current_ir_graph, block, x,
3379 new_r_Const_long(current_ir_graph, block, mode_Iu,
3380 get_mode_size_bits(mode) - 1),
3382 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3384 } else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) &&
3385 classify_Const(t) == CNST_ONE &&
3386 classify_Const(f) == CNST_NULL) {
3388 * Mux(x:T >/>= 0, 0, 1) -> Shr(-x, sizeof_bits(T) - 1)
3392 n = new_rd_Shr(get_irn_dbg_info(n),
3393 current_ir_graph, block,
3394 new_r_Minus(current_ir_graph, block, x, mode),
3395 new_r_Const_long(current_ir_graph, block, mode_Iu,
3396 get_mode_size_bits(mode) - 1),
3398 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3405 return arch_transform_node_Mux(n);
3406 } /* transform_node_Mux */
3409 * Optimize a Psi into some simpler cases.
3411 static ir_node *transform_node_Psi(ir_node *n) {
3413 return transform_node_Mux(n);
3416 } /* transform_node_Psi */
3419 * Tries several [inplace] [optimizing] transformations and returns an
3420 * equivalent node. The difference to equivalent_node() is that these
3421 * transformations _do_ generate new nodes, and thus the old node must
3422 * not be freed even if the equivalent node isn't the old one.
3424 static ir_node *transform_node(ir_node *n) {
3425 if (n->op->ops.transform_node)
3426 n = n->op->ops.transform_node(n);
3428 } /* transform_node */
3431 * Sets the default transform node operation for an ir_op_ops.
3433 * @param code the opcode for the default operation
3434 * @param ops the operations initialized
3439 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
3443 ops->transform_node = transform_node_##a; \
3476 } /* firm_set_default_transform_node */
3479 /* **************** Common Subexpression Elimination **************** */
3481 /** The size of the hash table used, should estimate the number of nodes
3483 #define N_IR_NODES 512
3485 /** Compares the attributes of two Const nodes. */
3486 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
3487 return (get_Const_tarval(a) != get_Const_tarval(b))
3488 || (get_Const_type(a) != get_Const_type(b));
3489 } /* node_cmp_attr_Const */
3491 /** Compares the attributes of two Proj nodes. */
3492 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
3493 return get_irn_proj_attr (a) != get_irn_proj_attr (b);
3494 } /* node_cmp_attr_Proj */
3496 /** Compares the attributes of two Filter nodes. */
3497 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
3498 return get_Filter_proj(a) != get_Filter_proj(b);
3499 } /* node_cmp_attr_Filter */
3501 /** Compares the attributes of two Alloc nodes. */
3502 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
3503 return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where)
3504 || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type);
3505 } /* node_cmp_attr_Alloc */
3507 /** Compares the attributes of two Free nodes. */
3508 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
3509 return (get_irn_free_attr(a).where != get_irn_free_attr(b).where)
3510 || (get_irn_free_attr(a).type != get_irn_free_attr(b).type);
3511 } /* node_cmp_attr_Free */
3513 /** Compares the attributes of two SymConst nodes. */
3514 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
3515 return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num)
3516 || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p)
3517 || (get_irn_symconst_attr(a).tp != get_irn_symconst_attr(b).tp);
3518 } /* node_cmp_attr_SymConst */
3520 /** Compares the attributes of two Call nodes. */
3521 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
3522 return (get_irn_call_attr(a) != get_irn_call_attr(b));
3523 } /* node_cmp_attr_Call */
3525 /** Compares the attributes of two Sel nodes. */
3526 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
3527 return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind)
3528 || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name)
3529 || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner)
3530 || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name)
3531 || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type);
3532 } /* node_cmp_attr_Sel */
3534 /** Compares the attributes of two Phi nodes. */
3535 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
3536 /* we can only enter this function if both nodes have the same number of inputs,
3537 hence it is enough to check if one of them is a Phi0 */
3539 /* check the Phi0 attribute */
3540 return get_irn_phi0_attr(a) != get_irn_phi0_attr(b);
3543 } /* node_cmp_attr_Phi */
3545 /** Compares the attributes of two Conv nodes. */
3546 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
3547 return get_Conv_strict(a) != get_Conv_strict(b);
3548 } /* node_cmp_attr_Conv */
3550 /** Compares the attributes of two Cast nodes. */
3551 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
3552 return get_Cast_type(a) != get_Cast_type(b);
3553 } /* node_cmp_attr_Cast */
3555 /** Compares the attributes of two Load nodes. */
3556 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
3557 if (get_Load_volatility(a) == volatility_is_volatile ||
3558 get_Load_volatility(b) == volatility_is_volatile)
3559 /* NEVER do CSE on volatile Loads */
3562 return get_Load_mode(a) != get_Load_mode(b);
3563 } /* node_cmp_attr_Load */
3565 /** Compares the attributes of two Store nodes. */
3566 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
3567 /* NEVER do CSE on volatile Stores */
3568 return (get_Store_volatility(a) == volatility_is_volatile ||
3569 get_Store_volatility(b) == volatility_is_volatile);
3570 } /* node_cmp_attr_Store */
3572 /** Compares the attributes of two Confirm nodes. */
3573 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
3574 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
3575 } /* node_cmp_attr_Confirm */
3577 /** Compares the attributes of two ASM nodes. */
3578 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
3580 const ir_asm_constraint *ca;
3581 const ir_asm_constraint *cb;
3584 if (get_ASM_text(a) != get_ASM_text(b));
3587 /* Should we really check the constraints here? Should be better, but is strange. */
3588 n = get_ASM_n_input_constraints(a);
3589 if (n != get_ASM_n_input_constraints(b))
3592 ca = get_ASM_input_constraints(a);
3593 cb = get_ASM_input_constraints(b);
3594 for (i = 0; i < n; ++i) {
3595 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
3599 n = get_ASM_n_output_constraints(a);
3600 if (n != get_ASM_n_output_constraints(b))
3603 ca = get_ASM_output_constraints(a);
3604 cb = get_ASM_output_constraints(b);
3605 for (i = 0; i < n; ++i) {
3606 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
3610 n = get_ASM_n_clobbers(a);
3611 if (n != get_ASM_n_clobbers(b))
3614 cla = get_ASM_clobbers(a);
3615 clb = get_ASM_clobbers(b);
3616 for (i = 0; i < n; ++i) {
3617 if (cla[i] != clb[i])
3621 } /* node_cmp_attr_ASM */
3624 * Set the default node attribute compare operation for an ir_op_ops.
3626 * @param code the opcode for the default operation
3627 * @param ops the operations initialized
3632 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
3636 ops->node_cmp_attr = node_cmp_attr_##a; \
3661 } /* firm_set_default_node_cmp_attr */
3664 * Compare function for two nodes in the hash table. Gets two
3665 * nodes as parameters. Returns 0 if the nodes are a cse.
3667 int identities_cmp(const void *elt, const void *key) {
3674 if (a == b) return 0;
3676 if ((get_irn_op(a) != get_irn_op(b)) ||
3677 (get_irn_mode(a) != get_irn_mode(b))) return 1;
3679 /* compare if a's in and b's in are of equal length */
3680 irn_arity_a = get_irn_intra_arity (a);
3681 if (irn_arity_a != get_irn_intra_arity(b))
3684 /* for block-local cse and op_pin_state_pinned nodes: */
3685 if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) {
3686 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
3690 /* compare a->in[0..ins] with b->in[0..ins] */
3691 for (i = 0; i < irn_arity_a; i++)
3692 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
3696 * here, we already now that the nodes are identical except their
3699 if (a->op->ops.node_cmp_attr)
3700 return a->op->ops.node_cmp_attr(a, b);
3703 } /* identities_cmp */
3706 * Calculate a hash value of a node.
3708 unsigned ir_node_hash(ir_node *node) {
3712 if (node->op == op_Const) {
3713 /* special value for const, as they only differ in their tarval. */
3714 h = HASH_PTR(node->attr.con.tv);
3715 h = 9*h + HASH_PTR(get_irn_mode(node));
3716 } else if (node->op == op_SymConst) {
3717 /* special value for const, as they only differ in their symbol. */
3718 h = HASH_PTR(node->attr.symc.sym.type_p);
3719 h = 9*h + HASH_PTR(get_irn_mode(node));
3722 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
3723 h = irn_arity = get_irn_intra_arity(node);
3725 /* consider all in nodes... except the block if not a control flow. */
3726 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
3727 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
3731 h = 9*h + HASH_PTR(get_irn_mode(node));
3733 h = 9*h + HASH_PTR(get_irn_op(node));
3737 } /* ir_node_hash */
3739 pset *new_identities(void) {
3740 return new_pset(identities_cmp, N_IR_NODES);
3741 } /* new_identities */
3743 void del_identities(pset *value_table) {
3744 del_pset(value_table);
3745 } /* del_identities */
3748 * Return the canonical node computing the same value as n.
3750 * @param value_table The value table
3751 * @param n The node to lookup
3753 * Looks up the node in a hash table.
3755 * For Const nodes this is performed in the constructor, too. Const
3756 * nodes are extremely time critical because of their frequent use in
3757 * constant string arrays.
3759 static INLINE ir_node *identify(pset *value_table, ir_node *n) {
3762 if (!value_table) return n;
3764 if (get_opt_reassociation()) {
3765 if (is_op_commutative(get_irn_op(n))) {
3766 ir_node *l = get_binop_left(n);
3767 ir_node *r = get_binop_right(n);
3769 /* for commutative operators perform a OP b == b OP a */
3770 if (get_irn_idx(l) > get_irn_idx(r)) {
3771 set_binop_left(n, r);
3772 set_binop_right(n, l);
3777 o = pset_find(value_table, n, ir_node_hash(n));
3786 * During construction we set the op_pin_state_pinned flag in the graph right when the
3787 * optimization is performed. The flag turning on procedure global cse could
3788 * be changed between two allocations. This way we are safe.
3790 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
3793 n = identify(value_table, n);
3794 if (get_irn_n(old, -1) != get_irn_n(n, -1))
3795 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3797 } /* identify_cons */
3800 * Return the canonical node computing the same value as n.
3801 * Looks up the node in a hash table, enters it in the table
3802 * if it isn't there yet.
3804 ir_node *identify_remember(pset *value_table, ir_node *n) {
3807 if (!value_table) return n;
3809 if (get_opt_reassociation()) {
3810 if (is_op_commutative(get_irn_op(n))) {
3811 ir_node *l = get_binop_left(n);
3812 ir_node *r = get_binop_right(n);
3813 int l_idx = get_irn_idx(l);
3814 int r_idx = get_irn_idx(r);
3816 /* For commutative operators perform a OP b == b OP a but keep
3817 constants on the RIGHT side. This helps greatly in some optimizations.
3818 Moreover we use the idx number to make the form deterministic. */
3819 if (is_irn_constlike(l))
3821 if (is_irn_constlike(r))
3823 if (l_idx < r_idx) {
3824 set_binop_left(n, r);
3825 set_binop_right(n, l);
3830 /* lookup or insert in hash table with given hash key. */
3831 o = pset_insert(value_table, n, ir_node_hash(n));
3838 } /* identify_remember */
3840 /* Add a node to the identities value table. */
3841 void add_identities(pset *value_table, ir_node *node) {
3842 if (get_opt_cse() && is_no_Block(node))
3843 identify_remember(value_table, node);
3844 } /* add_identities */
3846 /* Visit each node in the value table of a graph. */
3847 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
3849 ir_graph *rem = current_ir_graph;
3851 current_ir_graph = irg;
3852 foreach_pset(irg->value_table, node)
3854 current_ir_graph = rem;
3855 } /* visit_all_identities */
3858 * Garbage in, garbage out. If a node has a dead input, i.e., the
3859 * Bad node is input to the node, return the Bad node.
3861 static INLINE ir_node *gigo(ir_node *node) {
3863 ir_op *op = get_irn_op(node);
3865 /* remove garbage blocks by looking at control flow that leaves the block
3866 and replacing the control flow by Bad. */
3867 if (get_irn_mode(node) == mode_X) {
3868 ir_node *block = get_nodes_block(skip_Proj(node));
3870 /* Don't optimize nodes in immature blocks. */
3871 if (!get_Block_matured(block)) return node;
3872 /* Don't optimize End, may have Bads. */
3873 if (op == op_End) return node;
3875 if (is_Block(block)) {
3876 irn_arity = get_irn_arity(block);
3877 for (i = 0; i < irn_arity; i++) {
3878 if (!is_Bad(get_irn_n(block, i)))
3881 if (i == irn_arity) {
3882 ir_graph *irg = get_irn_irg(block);
3883 /* the start block is never dead */
3884 if (block != get_irg_start_block(irg)
3885 && block != get_irg_end_block(irg))
3891 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
3892 blocks predecessors is dead. */
3893 if (op != op_Block && op != op_Phi && op != op_Tuple) {
3894 irn_arity = get_irn_arity(node);
3897 * Beware: we can only read the block of a non-floating node.
3899 if (is_irn_pinned_in_irg(node) &&
3900 is_Block_dead(get_nodes_block(node)))
3903 for (i = 0; i < irn_arity; i++) {
3904 ir_node *pred = get_irn_n(node, i);
3909 /* Propagating Unknowns here seems to be a bad idea, because
3910 sometimes we need a node as a input and did not want that
3912 However, it might be useful to move this into a later phase
3913 (if you think that optimizing such code is useful). */
3914 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
3915 return new_Unknown(get_irn_mode(node));
3920 /* With this code we violate the agreement that local_optimize
3921 only leaves Bads in Block, Phi and Tuple nodes. */
3922 /* If Block has only Bads as predecessors it's garbage. */
3923 /* If Phi has only Bads as predecessors it's garbage. */
3924 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
3925 irn_arity = get_irn_arity(node);
3926 for (i = 0; i < irn_arity; i++) {
3927 if (!is_Bad(get_irn_n(node, i))) break;
3929 if (i == irn_arity) node = new_Bad();
3936 * These optimizations deallocate nodes from the obstack.
3937 * It can only be called if it is guaranteed that no other nodes
3938 * reference this one, i.e., right after construction of a node.
3940 * @param n The node to optimize
3942 * current_ir_graph must be set to the graph of the node!
3944 ir_node *optimize_node(ir_node *n) {
3947 ir_opcode iro = get_irn_opcode(n);
3949 /* Always optimize Phi nodes: part of the construction. */
3950 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
3952 /* constant expression evaluation / constant folding */
3953 if (get_opt_constant_folding()) {
3954 /* neither constants nor Tuple values can be evaluated */
3955 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
3956 /* try to evaluate */
3957 tv = computed_value(n);
3958 if (tv != tarval_bad) {
3960 ir_type *old_tp = get_irn_type(n);
3961 int i, arity = get_irn_arity(n);
3965 * Try to recover the type of the new expression.
3967 for (i = 0; i < arity && !old_tp; ++i)
3968 old_tp = get_irn_type(get_irn_n(n, i));
3971 * we MUST copy the node here temporary, because it's still needed
3972 * for DBG_OPT_CSTEVAL
3974 node_size = offsetof(ir_node, attr) + n->op->attr_size;
3975 oldn = alloca(node_size);
3977 memcpy(oldn, n, node_size);
3978 CLONE_ARR_A(ir_node *, oldn->in, n->in);
3980 /* ARG, copy the in array, we need it for statistics */
3981 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
3983 /* note the inplace edges module */
3984 edges_node_deleted(n, current_ir_graph);
3986 /* evaluation was successful -- replace the node. */
3987 irg_kill_node(current_ir_graph, n);
3988 nw = new_Const(get_tarval_mode (tv), tv);
3990 if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv))
3991 set_Const_type(nw, old_tp);
3992 DBG_OPT_CSTEVAL(oldn, nw);
3998 /* remove unnecessary nodes */
3999 if (get_opt_constant_folding() ||
4000 (iro == iro_Phi) || /* always optimize these nodes. */
4002 (iro == iro_Proj) ||
4003 (iro == iro_Block) ) /* Flags tested local. */
4004 n = equivalent_node(n);
4006 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
4008 /* Common Subexpression Elimination.
4010 * Checks whether n is already available.
4011 * The block input is used to distinguish different subexpressions. Right
4012 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
4013 * subexpressions within a block.
4016 n = identify_cons(current_ir_graph->value_table, n);
4019 edges_node_deleted(oldn, current_ir_graph);
4021 /* We found an existing, better node, so we can deallocate the old node. */
4022 irg_kill_node(current_ir_graph, oldn);
4026 /* Some more constant expression evaluation that does not allow to
4028 iro = get_irn_opcode(n);
4029 if (get_opt_constant_folding() ||
4030 (iro == iro_Cond) ||
4031 (iro == iro_Proj)) /* Flags tested local. */
4032 n = transform_node(n);
4034 /* Remove nodes with dead (Bad) input.
4035 Run always for transformation induced Bads. */
4038 /* Now we have a legal, useful node. Enter it in hash table for CSE */
4039 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
4040 n = identify_remember(current_ir_graph->value_table, n);
4044 } /* optimize_node */
4048 * These optimizations never deallocate nodes (in place). This can cause dead
4049 * nodes lying on the obstack. Remove these by a dead node elimination,
4050 * i.e., a copying garbage collection.
4052 ir_node *optimize_in_place_2(ir_node *n) {
4055 ir_opcode iro = get_irn_opcode(n);
4057 if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n;
4059 /* constant expression evaluation / constant folding */
4060 if (get_opt_constant_folding()) {
4061 /* neither constants nor Tuple values can be evaluated */
4062 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
4063 /* try to evaluate */
4064 tv = computed_value(n);
4065 if (tv != tarval_bad) {
4066 /* evaluation was successful -- replace the node. */
4067 ir_type *old_tp = get_irn_type(n);
4068 int i, arity = get_irn_arity(n);
4071 * Try to recover the type of the new expression.
4073 for (i = 0; i < arity && !old_tp; ++i)
4074 old_tp = get_irn_type(get_irn_n(n, i));
4076 n = new_Const(get_tarval_mode(tv), tv);
4078 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
4079 set_Const_type(n, old_tp);
4081 DBG_OPT_CSTEVAL(oldn, n);
4087 /* remove unnecessary nodes */
4088 if (get_opt_constant_folding() ||
4089 (iro == iro_Phi) || /* always optimize these nodes. */
4090 (iro == iro_Id) || /* ... */
4091 (iro == iro_Proj) || /* ... */
4092 (iro == iro_Block) ) /* Flags tested local. */
4093 n = equivalent_node(n);
4095 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
4097 /** common subexpression elimination **/
4098 /* Checks whether n is already available. */
4099 /* The block input is used to distinguish different subexpressions. Right
4100 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
4101 subexpressions within a block. */
4102 if (get_opt_cse()) {
4103 n = identify(current_ir_graph->value_table, n);
4106 /* Some more constant expression evaluation. */
4107 iro = get_irn_opcode(n);
4108 if (get_opt_constant_folding() ||
4109 (iro == iro_Cond) ||
4110 (iro == iro_Proj)) /* Flags tested local. */
4111 n = transform_node(n);
4113 /* Remove nodes with dead (Bad) input.
4114 Run always for transformation induced Bads. */
4117 /* Now we can verify the node, as it has no dead inputs any more. */
4120 /* Now we have a legal, useful node. Enter it in hash table for cse.
4121 Blocks should be unique anyways. (Except the successor of start:
4122 is cse with the start block!) */
4123 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
4124 n = identify_remember(current_ir_graph->value_table, n);
4127 } /* optimize_in_place_2 */
4130 * Wrapper for external use, set proper status bits after optimization.
4132 ir_node *optimize_in_place(ir_node *n) {
4133 /* Handle graph state */
4134 assert(get_irg_phase_state(current_ir_graph) != phase_building);
4136 if (get_opt_global_cse())
4137 set_irg_pinned(current_ir_graph, op_pin_state_floats);
4138 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
4139 set_irg_outs_inconsistent(current_ir_graph);
4141 /* FIXME: Maybe we could also test whether optimizing the node can
4142 change the control graph. */
4143 set_irg_doms_inconsistent(current_ir_graph);
4144 return optimize_in_place_2(n);
4145 } /* optimize_in_place */
4148 * Sets the default operation for an ir_ops.
4150 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
4151 ops = firm_set_default_computed_value(code, ops);
4152 ops = firm_set_default_equivalent_node(code, ops);
4153 ops = firm_set_default_transform_node(code, ops);
4154 ops = firm_set_default_node_cmp_attr(code, ops);
4155 ops = firm_set_default_get_type(code, ops);
4156 ops = firm_set_default_get_type_attr(code, ops);
4157 ops = firm_set_default_get_entity_attr(code, ops);
4160 } /* firm_set_default_operations */