2 * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief iropt --- optimizations intertwined with IR construction.
23 * @author Christian Schaefer, Goetz Lindenmaier, Michael Beck
33 #include "irgraph_t.h"
34 #include "iredges_t.h"
41 #include "dbginfo_t.h"
42 #include "iropt_dbg.h"
48 #include "opt_confirms.h"
49 #include "opt_polymorphy.h"
53 /* Make types visible to allow most efficient access */
57 * Return the value of a Constant.
59 static tarval *computed_value_Const(ir_node *n) {
60 return get_Const_tarval(n);
61 } /* computed_value_Const */
64 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
66 static tarval *computed_value_SymConst(ir_node *n) {
70 switch (get_SymConst_kind(n)) {
71 case symconst_type_size:
72 type = get_SymConst_type(n);
73 if (get_type_state(type) == layout_fixed)
74 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
76 case symconst_type_align:
77 type = get_SymConst_type(n);
78 if (get_type_state(type) == layout_fixed)
79 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
81 case symconst_ofs_ent:
82 ent = get_SymConst_entity(n);
83 type = get_entity_owner(ent);
84 if (get_type_state(type) == layout_fixed)
85 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
91 } /* computed_value_SymConst */
94 * Return the value of an Add.
96 static tarval *computed_value_Add(ir_node *n) {
97 ir_node *a = get_Add_left(n);
98 ir_node *b = get_Add_right(n);
100 tarval *ta = value_of(a);
101 tarval *tb = value_of(b);
103 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b)))
104 return tarval_add(ta, tb);
107 } /* computed_value_Add */
110 * Return the value of a Sub.
111 * Special case: a - a
113 static tarval *computed_value_Sub(ir_node *n) {
114 ir_node *a = get_Sub_left(n);
115 ir_node *b = get_Sub_right(n);
120 if (a == b && !is_Bad(a))
121 return get_mode_null(get_irn_mode(n));
126 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b)))
127 return tarval_sub(ta, tb);
130 } /* computed_value_Sub */
133 * Return the value of a Carry.
134 * Special : a op 0, 0 op b
136 static tarval *computed_value_Carry(ir_node *n) {
137 ir_node *a = get_binop_left(n);
138 ir_node *b = get_binop_right(n);
139 ir_mode *m = get_irn_mode(n);
141 tarval *ta = value_of(a);
142 tarval *tb = value_of(b);
144 if ((ta != tarval_bad) && (tb != tarval_bad)) {
146 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
148 if ( (classify_tarval(ta) == TV_CLASSIFY_NULL)
149 || (classify_tarval(tb) == TV_CLASSIFY_NULL))
150 return get_mode_null(m);
153 } /* computed_value_Carry */
156 * Return the value of a Borrow.
159 static tarval *computed_value_Borrow(ir_node *n) {
160 ir_node *a = get_binop_left(n);
161 ir_node *b = get_binop_right(n);
162 ir_mode *m = get_irn_mode(n);
164 tarval *ta = value_of(a);
165 tarval *tb = value_of(b);
167 if ((ta != tarval_bad) && (tb != tarval_bad)) {
168 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
169 } else if (classify_tarval(ta) == TV_CLASSIFY_NULL) {
170 return get_mode_null(m);
173 } /* computed_value_Borrow */
176 * Return the value of an unary Minus.
178 static tarval *computed_value_Minus(ir_node *n) {
179 ir_node *a = get_Minus_op(n);
180 tarval *ta = value_of(a);
182 if ((ta != tarval_bad) && mode_is_signed(get_irn_mode(a)))
183 return tarval_neg(ta);
186 } /* computed_value_Minus */
189 * Return the value of a Mul.
191 static tarval *computed_value_Mul(ir_node *n) {
192 ir_node *a = get_Mul_left(n);
193 ir_node *b = get_Mul_right(n);
195 tarval *ta = value_of(a);
196 tarval *tb = value_of(b);
198 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
199 return tarval_mul(ta, tb);
201 /* a*0 = 0 or 0*b = 0:
202 calls computed_value recursive and returns the 0 with proper
204 if ((ta != tarval_bad) && (ta == get_mode_null(get_tarval_mode(ta))))
206 if ((tb != tarval_bad) && (tb == get_mode_null(get_tarval_mode(tb))))
210 } /* computed_value_Mul */
213 * Return the value of a floating point Quot.
215 static tarval *computed_value_Quot(ir_node *n) {
216 ir_node *a = get_Quot_left(n);
217 ir_node *b = get_Quot_right(n);
219 tarval *ta = value_of(a);
220 tarval *tb = value_of(b);
222 /* This was missing in original implementation. Why? */
223 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
224 if (tb != get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */
225 return tarval_quo(ta, tb);
228 } /* computed_value_Quot */
231 * Calculate the value of an integer Div of two nodes.
232 * Special case: 0 / b
234 static tarval *do_computed_value_Div(ir_node *a, ir_node *b) {
235 tarval *ta = value_of(a);
236 tarval *tb = value_of(b);
238 /* Compute c1 / c2 or 0 / a, a != 0 */
239 if (ta != tarval_bad) {
240 if ((tb != tarval_bad) && (tb != get_mode_null(get_irn_mode(b)))) /* div by zero: return tarval_bad */
241 return tarval_div(ta, tb);
242 else if (ta == get_mode_null(get_tarval_mode(ta))) /* 0 / b == 0 */
246 } /* do_computed_value_Div */
249 * Return the value of an integer Div.
251 static tarval *computed_value_Div(ir_node *n) {
252 return do_computed_value_Div(get_Div_left(n), get_Div_right(n));
253 } /* computed_value_Div */
256 * Calculate the value of an integer Mod of two nodes.
257 * Special case: a % 1
259 static tarval *do_computed_value_Mod(ir_node *a, ir_node *b) {
260 tarval *ta = value_of(a);
261 tarval *tb = value_of(b);
263 /* Compute c1 % c2 or a % 1 */
264 if (tb != tarval_bad) {
265 if ((ta != tarval_bad) && (tb != get_mode_null(get_tarval_mode(tb)))) /* div by zero: return tarval_bad */
266 return tarval_mod(ta, tb);
267 else if (tb == get_mode_one(get_tarval_mode(tb))) /* x mod 1 == 0 */
268 return get_mode_null(get_irn_mode(a));
271 } /* do_computed_value_Mod */
274 * Return the value of an integer Mod.
276 static tarval *computed_value_Mod(ir_node *n) {
277 return do_computed_value_Mod(get_Mod_left(n), get_Mod_right(n));
278 } /* computed_value_Mod */
281 * Return the value of an Abs.
283 static tarval *computed_value_Abs(ir_node *n) {
284 ir_node *a = get_Abs_op(n);
285 tarval *ta = value_of(a);
287 if (ta != tarval_bad)
288 return tarval_abs(ta);
291 } /* computed_value_Abs */
294 * Return the value of an And.
295 * Special case: a & 0, 0 & b
297 static tarval *computed_value_And(ir_node *n) {
298 ir_node *a = get_And_left(n);
299 ir_node *b = get_And_right(n);
301 tarval *ta = value_of(a);
302 tarval *tb = value_of(b);
304 if ((ta != tarval_bad) && (tb != tarval_bad)) {
305 return tarval_and (ta, tb);
309 if ( (classify_tarval ((v = ta)) == TV_CLASSIFY_NULL)
310 || (classify_tarval ((v = tb)) == TV_CLASSIFY_NULL)) {
315 } /* computed_value_And */
318 * Return the value of an Or.
319 * Special case: a | 1...1, 1...1 | b
321 static tarval *computed_value_Or(ir_node *n) {
322 ir_node *a = get_Or_left(n);
323 ir_node *b = get_Or_right(n);
325 tarval *ta = value_of(a);
326 tarval *tb = value_of(b);
328 if ((ta != tarval_bad) && (tb != tarval_bad)) {
329 return tarval_or (ta, tb);
332 if ( (classify_tarval ((v = ta)) == TV_CLASSIFY_ALL_ONE)
333 || (classify_tarval ((v = tb)) == TV_CLASSIFY_ALL_ONE)) {
338 } /* computed_value_Or */
341 * Return the value of an Eor.
343 static tarval *computed_value_Eor(ir_node *n) {
344 ir_node *a = get_Eor_left(n);
345 ir_node *b = get_Eor_right(n);
350 return get_mode_null(get_irn_mode(n));
355 if ((ta != tarval_bad) && (tb != tarval_bad)) {
356 return tarval_eor (ta, tb);
359 } /* computed_value_Eor */
362 * Return the value of a Not.
364 static tarval *computed_value_Not(ir_node *n) {
365 ir_node *a = get_Not_op(n);
366 tarval *ta = value_of(a);
368 if (ta != tarval_bad)
369 return tarval_not(ta);
372 } /* computed_value_Not */
375 * Return the value of a Shl.
377 static tarval *computed_value_Shl(ir_node *n) {
378 ir_node *a = get_Shl_left(n);
379 ir_node *b = get_Shl_right(n);
381 tarval *ta = value_of(a);
382 tarval *tb = value_of(b);
384 if ((ta != tarval_bad) && (tb != tarval_bad)) {
385 return tarval_shl (ta, tb);
388 } /* computed_value_Shl */
391 * Return the value of a Shr.
393 static tarval *computed_value_Shr(ir_node *n) {
394 ir_node *a = get_Shr_left(n);
395 ir_node *b = get_Shr_right(n);
397 tarval *ta = value_of(a);
398 tarval *tb = value_of(b);
400 if ((ta != tarval_bad) && (tb != tarval_bad)) {
401 return tarval_shr (ta, tb);
404 } /* computed_value_Shr */
407 * Return the value of a Shrs.
409 static tarval *computed_value_Shrs(ir_node *n) {
410 ir_node *a = get_Shrs_left(n);
411 ir_node *b = get_Shrs_right(n);
413 tarval *ta = value_of(a);
414 tarval *tb = value_of(b);
416 if ((ta != tarval_bad) && (tb != tarval_bad)) {
417 return tarval_shrs (ta, tb);
420 } /* computed_value_Shrs */
423 * Return the value of a Rot.
425 static tarval *computed_value_Rot(ir_node *n) {
426 ir_node *a = get_Rot_left(n);
427 ir_node *b = get_Rot_right(n);
429 tarval *ta = value_of(a);
430 tarval *tb = value_of(b);
432 if ((ta != tarval_bad) && (tb != tarval_bad)) {
433 return tarval_rot (ta, tb);
436 } /* computed_value_Rot */
439 * Return the value of a Conv.
441 static tarval *computed_value_Conv(ir_node *n) {
442 ir_node *a = get_Conv_op(n);
443 tarval *ta = value_of(a);
445 if (ta != tarval_bad)
446 return tarval_convert_to(ta, get_irn_mode(n));
449 } /* computed_value_Conv */
452 * Return the value of a Proj(Cmp).
454 * This performs a first step of unreachable code elimination.
455 * Proj can not be computed, but folding a Cmp above the Proj here is
456 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
458 * There are several case where we can evaluate a Cmp node, see later.
460 static tarval *computed_value_Proj_Cmp(ir_node *n) {
461 ir_node *a = get_Proj_pred(n);
462 ir_node *aa = get_Cmp_left(a);
463 ir_node *ab = get_Cmp_right(a);
464 long proj_nr = get_Proj_proj(n);
467 * BEWARE: a == a is NOT always True for floating Point values, as
468 * NaN != NaN is defined, so we must check this here.
471 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
474 /* This is a trick with the bits used for encoding the Cmp
475 Proj numbers, the following statement is not the same:
476 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
477 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
480 tarval *taa = value_of(aa);
481 tarval *tab = value_of(ab);
482 ir_mode *mode = get_irn_mode(aa);
485 * The predecessors of Cmp are target values. We can evaluate
488 if ((taa != tarval_bad) && (tab != tarval_bad)) {
489 /* strange checks... */
490 pn_Cmp flags = tarval_cmp(taa, tab);
491 if (flags != pn_Cmp_False) {
492 return new_tarval_from_long (proj_nr & flags, mode_b);
495 /* for integer values, we can check against MIN/MAX */
496 else if (mode_is_int(mode)) {
497 /* MIN <=/> x. This results in true/false. */
498 if (taa == get_mode_min(mode)) {
499 /* a compare with the MIN value */
500 if (proj_nr == pn_Cmp_Le)
501 return get_tarval_b_true();
502 else if (proj_nr == pn_Cmp_Gt)
503 return get_tarval_b_false();
505 /* x >=/< MIN. This results in true/false. */
507 if (tab == get_mode_min(mode)) {
508 /* a compare with the MIN value */
509 if (proj_nr == pn_Cmp_Ge)
510 return get_tarval_b_true();
511 else if (proj_nr == pn_Cmp_Lt)
512 return get_tarval_b_false();
514 /* MAX >=/< x. This results in true/false. */
515 else if (taa == get_mode_max(mode)) {
516 if (proj_nr == pn_Cmp_Ge)
517 return get_tarval_b_true();
518 else if (proj_nr == pn_Cmp_Lt)
519 return get_tarval_b_false();
521 /* x <=/> MAX. This results in true/false. */
522 else if (tab == get_mode_max(mode)) {
523 if (proj_nr == pn_Cmp_Le)
524 return get_tarval_b_true();
525 else if (proj_nr == pn_Cmp_Gt)
526 return get_tarval_b_false();
530 * The predecessors are Allocs or (void*)(0) constants. Allocs never
531 * return NULL, they raise an exception. Therefore we can predict
535 ir_node *aaa = skip_Id(skip_Proj(aa));
536 ir_node *aba = skip_Id(skip_Proj(ab));
538 if ( ( (/* aa is ProjP and aaa is Alloc */
539 (get_irn_op(aa) == op_Proj)
540 && (mode_is_reference(get_irn_mode(aa)))
541 && (get_irn_op(aaa) == op_Alloc))
542 && ( (/* ab is NULL */
543 (get_irn_op(ab) == op_Const)
544 && (mode_is_reference(get_irn_mode(ab)))
545 && (get_Const_tarval(ab) == get_mode_null(get_irn_mode(ab))))
546 || (/* ab is other Alloc */
547 (get_irn_op(ab) == op_Proj)
548 && (mode_is_reference(get_irn_mode(ab)))
549 && (get_irn_op(aba) == op_Alloc)
551 || (/* aa is NULL and aba is Alloc */
552 (get_irn_op(aa) == op_Const)
553 && (mode_is_reference(get_irn_mode(aa)))
554 && (get_Const_tarval(aa) == get_mode_null(get_irn_mode(aa)))
555 && (get_irn_op(ab) == op_Proj)
556 && (mode_is_reference(get_irn_mode(ab)))
557 && (get_irn_op(aba) == op_Alloc)))
559 return new_tarval_from_long(proj_nr & pn_Cmp_Ne, mode_b);
562 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
563 } /* computed_value_Proj_Cmp */
566 * Return the value of a Proj, handle Proj(Cmp), Proj(Div), Proj(Mod),
567 * Proj(DivMod) and Proj(Quot).
569 static tarval *computed_value_Proj(ir_node *n) {
570 ir_node *a = get_Proj_pred(n);
573 switch (get_irn_opcode(a)) {
575 return computed_value_Proj_Cmp(n);
578 /* compute either the Div or the Mod part */
579 proj_nr = get_Proj_proj(n);
580 if (proj_nr == pn_DivMod_res_div)
581 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
582 else if (proj_nr == pn_DivMod_res_mod)
583 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
587 if (get_Proj_proj(n) == pn_Div_res)
588 return computed_value(a);
592 if (get_Proj_proj(n) == pn_Mod_res)
593 return computed_value(a);
597 if (get_Proj_proj(n) == pn_Quot_res)
598 return computed_value(a);
605 } /* computed_value_Proj */
608 * Calculate the value of a Mux: can be evaluated, if the
609 * sel and the right input are known.
611 static tarval *computed_value_Mux(ir_node *n) {
612 ir_node *sel = get_Mux_sel(n);
613 tarval *ts = value_of(sel);
615 if (ts == get_tarval_b_true()) {
616 ir_node *v = get_Mux_true(n);
619 else if (ts == get_tarval_b_false()) {
620 ir_node *v = get_Mux_false(n);
624 } /* computed_value_Mux */
627 * Calculate the value of a Psi: can be evaluated, if a condition is true
628 * and all previous conditions are false. If all conditions are false
629 * we evaluate to the default one.
631 static tarval *computed_value_Psi(ir_node *n) {
633 return computed_value_Mux(n);
635 } /* computed_value_Psi */
638 * Calculate the value of a Confirm: can be evaluated,
639 * if it has the form Confirm(x, '=', Const).
641 static tarval *computed_value_Confirm(ir_node *n) {
642 return get_Confirm_cmp(n) == pn_Cmp_Eq ?
643 value_of(get_Confirm_bound(n)) : tarval_bad;
644 } /* computed_value_Confirm */
647 * If the parameter n can be computed, return its value, else tarval_bad.
648 * Performs constant folding.
650 * @param n The node this should be evaluated
652 tarval *computed_value(ir_node *n) {
653 if (n->op->ops.computed_value)
654 return n->op->ops.computed_value(n);
656 } /* computed_value */
659 * Set the default computed_value evaluator in an ir_op_ops.
661 * @param code the opcode for the default operation
662 * @param ops the operations initialized
667 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
671 ops->computed_value = computed_value_##a; \
706 } /* firm_set_default_computed_value */
709 * Returns a equivalent block for another block.
710 * If the block has only one predecessor, this is
711 * the equivalent one. If the only predecessor of a block is
712 * the block itself, this is a dead block.
714 * If both predecessors of a block are the branches of a binary
715 * Cond, the equivalent block is Cond's block.
717 * If all predecessors of a block are bad or lies in a dead
718 * block, the current block is dead as well.
720 * Note, that blocks are NEVER turned into Bad's, instead
721 * the dead_block flag is set. So, never test for is_Bad(block),
722 * always use is_dead_Block(block).
724 static ir_node *equivalent_node_Block(ir_node *n)
727 int n_preds = get_Block_n_cfgpreds(n);
729 /* The Block constructor does not call optimize, but mature_immBlock
730 calls the optimization. */
731 assert(get_Block_matured(n));
733 /* Straightening: a single entry Block following a single exit Block
734 can be merged, if it is not the Start block. */
735 /* !!! Beware, all Phi-nodes of n must have been optimized away.
736 This should be true, as the block is matured before optimize is called.
737 But what about Phi-cycles with the Phi0/Id that could not be resolved?
738 Remaining Phi nodes are just Ids. */
739 if ((n_preds == 1) && (get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp)) {
740 ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0));
741 if (predblock == oldn) {
742 /* Jmp jumps into the block it is in -- deal self cycle. */
743 n = set_Block_dead(n);
744 DBG_OPT_DEAD_BLOCK(oldn, n);
745 } else if (get_opt_control_flow_straightening()) {
747 DBG_OPT_STG(oldn, n);
749 } else if ((n_preds == 1) &&
750 (get_irn_op(skip_Proj(get_Block_cfgpred(n, 0))) == op_Cond)) {
751 ir_node *predblock = get_Block_cfgpred_block(n, 0);
752 if (predblock == oldn) {
753 /* Jmp jumps into the block it is in -- deal self cycle. */
754 n = set_Block_dead(n);
755 DBG_OPT_DEAD_BLOCK(oldn, n);
757 } else if ((n_preds == 2) &&
758 (get_opt_control_flow_weak_simplification())) {
759 /* Test whether Cond jumps twice to this block
760 * The more general case which more than 2 predecessors is handles
761 * in optimize_cf(), we handle only this special case for speed here.
763 ir_node *a = get_Block_cfgpred(n, 0);
764 ir_node *b = get_Block_cfgpred(n, 1);
766 if ((get_irn_op(a) == op_Proj) &&
767 (get_irn_op(b) == op_Proj) &&
768 (get_Proj_pred(a) == get_Proj_pred(b)) &&
769 (get_irn_op(get_Proj_pred(a)) == op_Cond) &&
770 (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) {
771 /* Also a single entry Block following a single exit Block. Phis have
772 twice the same operand and will be optimized away. */
773 n = get_nodes_block(get_Proj_pred(a));
774 DBG_OPT_IFSIM1(oldn, a, b, n);
776 } else if (get_opt_unreachable_code() &&
777 (n != get_irg_start_block(current_ir_graph)) &&
778 (n != get_irg_end_block(current_ir_graph)) ) {
781 /* If all inputs are dead, this block is dead too, except if it is
782 the start or end block. This is one step of unreachable code
784 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
785 ir_node *pred = get_Block_cfgpred(n, i);
788 if (is_Bad(pred)) continue;
789 pred_blk = get_nodes_block(skip_Proj(pred));
791 if (is_Block_dead(pred_blk)) continue;
794 /* really found a living input */
799 n = set_Block_dead(n);
800 DBG_OPT_DEAD_BLOCK(oldn, n);
805 } /* equivalent_node_Block */
808 * Returns a equivalent node for a Jmp, a Bad :-)
809 * Of course this only happens if the Block of the Jmp is dead.
811 static ir_node *equivalent_node_Jmp(ir_node *n) {
812 /* unreachable code elimination */
813 if (is_Block_dead(get_nodes_block(n)))
817 } /* equivalent_node_Jmp */
819 /** Raise is handled in the same way as Jmp. */
820 #define equivalent_node_Raise equivalent_node_Jmp
823 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
824 See transform_node_Proj_Cond(). */
827 * Optimize operations that are commutative and have neutral 0,
828 * so a op 0 = 0 op a = a.
830 static ir_node *equivalent_node_neutral_zero(ir_node *n)
834 ir_node *a = get_binop_left(n);
835 ir_node *b = get_binop_right(n);
840 /* After running compute_node there is only one constant predecessor.
841 Find this predecessors value and remember the other node: */
842 if ((tv = value_of(a)) != tarval_bad) {
844 } else if ((tv = value_of(b)) != tarval_bad) {
849 /* If this predecessors constant value is zero, the operation is
850 * unnecessary. Remove it.
852 * Beware: If n is a Add, the mode of on and n might be different
853 * which happens in this rare construction: NULL + 3.
854 * Then, a Conv would be needed which we cannot include here.
856 if (classify_tarval (tv) == TV_CLASSIFY_NULL) {
857 if (get_irn_mode(on) == get_irn_mode(n)) {
860 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
865 } /* equivalent_node_neutral_zero */
868 * Eor is commutative and has neutral 0.
870 #define equivalent_node_Eor equivalent_node_neutral_zero
873 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
875 * The second one looks strange, but this construct
876 * is used heavily in the LCC sources :-).
878 * Beware: The Mode of an Add may be different than the mode of its
879 * predecessors, so we could not return a predecessors in all cases.
881 static ir_node *equivalent_node_Add(ir_node *n) {
883 ir_node *left, *right;
884 ir_mode *mode = get_irn_mode(n);
886 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
887 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
890 n = equivalent_node_neutral_zero(n);
894 left = get_Add_left(n);
895 right = get_Add_right(n);
897 if (get_irn_op(left) == op_Sub) {
898 if (get_Sub_right(left) == right) {
901 n = get_Sub_left(left);
902 if (mode == get_irn_mode(n)) {
903 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
908 if (get_irn_op(right) == op_Sub) {
909 if (get_Sub_right(right) == left) {
912 n = get_Sub_left(right);
913 if (mode == get_irn_mode(n)) {
914 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
920 } /* equivalent_node_Add */
923 * optimize operations that are not commutative but have neutral 0 on left,
926 static ir_node *equivalent_node_left_zero(ir_node *n) {
929 ir_node *a = get_binop_left(n);
930 ir_node *b = get_binop_right(n);
932 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
935 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
938 } /* equivalent_node_left_zero */
940 #define equivalent_node_Shl equivalent_node_left_zero
941 #define equivalent_node_Shr equivalent_node_left_zero
942 #define equivalent_node_Shrs equivalent_node_left_zero
943 #define equivalent_node_Rot equivalent_node_left_zero
946 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
948 * The second one looks strange, but this construct
949 * is used heavily in the LCC sources :-).
951 * Beware: The Mode of a Sub may be different than the mode of its
952 * predecessors, so we could not return a predecessors in all cases.
954 static ir_node *equivalent_node_Sub(ir_node *n) {
957 ir_mode *mode = get_irn_mode(n);
959 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
960 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
964 b = get_Sub_right(n);
966 /* Beware: modes might be different */
967 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
968 if (mode == get_irn_mode(a)) {
971 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
973 } else if (get_irn_op(a) == op_Add) {
974 if (mode_wrap_around(mode)) {
975 ir_node *left = get_Add_left(a);
976 ir_node *right = get_Add_right(a);
979 if (mode == get_irn_mode(right)) {
981 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
983 } else if (right == b) {
984 if (mode == get_irn_mode(left)) {
986 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
992 } /* equivalent_node_Sub */
996 * Optimize an "idempotent unary op", ie op(op(n)) = n.
999 * -(-a) == a, but might overflow two times.
1000 * We handle it anyway here but the better way would be a
1001 * flag. This would be needed for Pascal for instance.
1003 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
1005 ir_node *pred = get_unop_op(n);
1007 /* optimize symmetric unop */
1008 if (get_irn_op(pred) == get_irn_op(n)) {
1009 n = get_unop_op(pred);
1010 DBG_OPT_ALGSIM2(oldn, pred, n);
1013 } /* equivalent_node_idempotent_unop */
1015 /** Optimize Not(Not(x)) == x. */
1016 #define equivalent_node_Not equivalent_node_idempotent_unop
1018 /** --x == x ??? Is this possible or can --x raise an
1019 out of bounds exception if min =! max? */
1020 #define equivalent_node_Minus equivalent_node_idempotent_unop
1023 * Optimize a * 1 = 1 * a = a.
1025 static ir_node *equivalent_node_Mul(ir_node *n) {
1027 ir_node *a = get_Mul_left(n);
1028 ir_node *b = get_Mul_right(n);
1030 /* Mul is commutative and has again an other neutral element. */
1031 if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) {
1033 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1034 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) {
1036 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1039 } /* equivalent_node_Mul */
1042 * Optimize a / 1 = a.
1044 static ir_node *equivalent_node_Div(ir_node *n) {
1045 ir_node *a = get_Div_left(n);
1046 ir_node *b = get_Div_right(n);
1048 /* Div is not commutative. */
1049 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1050 /* Turn Div into a tuple (mem, bad, a) */
1051 ir_node *mem = get_Div_mem(n);
1052 ir_node *blk = get_nodes_block(n);
1053 turn_into_tuple(n, pn_Div_max);
1054 set_Tuple_pred(n, pn_Div_M, mem);
1055 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
1056 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
1057 set_Tuple_pred(n, pn_Div_res, a);
1060 } /* equivalent_node_Div */
1063 * Optimize a / 1.0 = a.
1065 static ir_node *equivalent_node_Quot(ir_node *n) {
1066 ir_node *a = get_Quot_left(n);
1067 ir_node *b = get_Quot_right(n);
1069 /* Div is not commutative. */
1070 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* Quot(x, 1) == x */
1071 /* Turn Quot into a tuple (mem, jmp, bad, a) */
1072 ir_node *mem = get_Quot_mem(n);
1073 ir_node *blk = get_nodes_block(n);
1074 turn_into_tuple(n, pn_Quot_max);
1075 set_Tuple_pred(n, pn_Quot_M, mem);
1076 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
1077 set_Tuple_pred(n, pn_Quot_X_except, new_Bad()); /* no exception */
1078 set_Tuple_pred(n, pn_Quot_res, a);
1081 } /* equivalent_node_Quot */
1084 * Optimize a / 1 = a.
1086 static ir_node *equivalent_node_DivMod(ir_node *n) {
1087 ir_node *b = get_DivMod_right(n);
1089 /* Div is not commutative. */
1090 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1091 /* Turn DivMod into a tuple (mem, jmp, bad, a, 0) */
1092 ir_node *a = get_DivMod_left(n);
1093 ir_node *mem = get_Div_mem(n);
1094 ir_node *blk = get_nodes_block(n);
1095 ir_mode *mode = get_DivMod_resmode(n);
1097 turn_into_tuple(n, pn_DivMod_max);
1098 set_Tuple_pred(n, pn_DivMod_M, mem);
1099 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
1100 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1101 set_Tuple_pred(n, pn_DivMod_res_div, a);
1102 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1105 } /* equivalent_node_DivMod */
1108 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1110 static ir_node *equivalent_node_Or(ir_node *n) {
1113 ir_node *a = get_Or_left(n);
1114 ir_node *b = get_Or_right(n);
1117 n = a; /* Or has it's own neutral element */
1118 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1119 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) {
1121 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1122 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
1124 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1128 } /* equivalent_node_Or */
1131 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a.
1133 static ir_node *equivalent_node_And(ir_node *n) {
1136 ir_node *a = get_And_left(n);
1137 ir_node *b = get_And_right(n);
1140 n = a; /* And has it's own neutral element */
1141 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1142 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) {
1144 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1145 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) {
1147 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1150 } /* equivalent_node_And */
1153 * Try to remove useless Conv's:
1155 static ir_node *equivalent_node_Conv(ir_node *n) {
1157 ir_node *a = get_Conv_op(n);
1160 ir_mode *n_mode = get_irn_mode(n);
1161 ir_mode *a_mode = get_irn_mode(a);
1163 if (n_mode == a_mode) { /* No Conv necessary */
1164 /* leave strict floating point Conv's */
1165 if (get_Conv_strict(n))
1168 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1169 } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */
1173 n_mode = get_irn_mode(n);
1174 b_mode = get_irn_mode(b);
1176 if (n_mode == b_mode) {
1177 if (n_mode == mode_b) {
1178 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1179 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1180 } else if (mode_is_int(n_mode) || mode_is_character(n_mode)) {
1181 if (smaller_mode(b_mode, a_mode)){
1182 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1183 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1189 } /* equivalent_node_Conv */
1192 * A Cast may be removed if the type of the previous node
1193 * is already the type of the Cast.
1195 static ir_node *equivalent_node_Cast(ir_node *n) {
1197 ir_node *pred = get_Cast_op(n);
1199 if (get_irn_type(pred) == get_Cast_type(n)) {
1201 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1204 } /* equivalent_node_Cast */
1207 * Several optimizations:
1208 * - no Phi in start block.
1209 * - remove Id operators that are inputs to Phi
1210 * - fold Phi-nodes, iff they have only one predecessor except
1213 static ir_node *equivalent_node_Phi(ir_node *n) {
1217 ir_node *block = NULL; /* to shutup gcc */
1218 ir_node *first_val = NULL; /* to shutup gcc */
1220 if (!get_opt_normalize()) return n;
1222 n_preds = get_Phi_n_preds(n);
1224 block = get_nodes_block(n);
1225 /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
1226 assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
1227 if ((is_Block_dead(block)) || /* Control dead */
1228 (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */
1229 return new_Bad(); /* in the Start Block. */
1231 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1233 /* If the Block has a Bad pred, we also have one. */
1234 for (i = 0; i < n_preds; ++i)
1235 if (is_Bad(get_Block_cfgpred(block, i)))
1236 set_Phi_pred(n, i, new_Bad());
1238 /* Find first non-self-referencing input */
1239 for (i = 0; i < n_preds; ++i) {
1240 first_val = get_Phi_pred(n, i);
1241 if ( (first_val != n) /* not self pointer */
1243 && (! is_Bad(first_val))
1245 ) { /* value not dead */
1246 break; /* then found first value. */
1251 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1255 /* search for rest of inputs, determine if any of these
1256 are non-self-referencing */
1257 while (++i < n_preds) {
1258 ir_node *scnd_val = get_Phi_pred(n, i);
1259 if ( (scnd_val != n)
1260 && (scnd_val != first_val)
1262 && (! is_Bad(scnd_val))
1270 /* Fold, if no multiple distinct non-self-referencing inputs */
1272 DBG_OPT_PHI(oldn, n);
1275 } /* equivalent_node_Phi */
1278 * Several optimizations:
1279 * - no Sync in start block.
1280 * - fold Sync-nodes, iff they have only one predecessor except
1283 static ir_node *equivalent_node_Sync(ir_node *n) {
1287 ir_node *first_val = NULL; /* to shutup gcc */
1289 if (!get_opt_normalize()) return n;
1291 n_preds = get_Sync_n_preds(n);
1293 /* Find first non-self-referencing input */
1294 for (i = 0; i < n_preds; ++i) {
1295 first_val = get_Sync_pred(n, i);
1296 if ((first_val != n) /* not self pointer */ &&
1297 (! is_Bad(first_val))
1298 ) { /* value not dead */
1299 break; /* then found first value. */
1304 /* A totally Bad or self-referencing Sync (we didn't break the above loop) */
1307 /* search the rest of inputs, determine if any of these
1308 are non-self-referencing */
1309 while (++i < n_preds) {
1310 ir_node *scnd_val = get_Sync_pred(n, i);
1311 if ((scnd_val != n) &&
1312 (scnd_val != first_val) &&
1313 (! is_Bad(scnd_val))
1319 /* Fold, if no multiple distinct non-self-referencing inputs */
1321 DBG_OPT_SYNC(oldn, n);
1324 } /* equivalent_node_Sync */
1327 * Optimize Proj(Tuple) and gigo() for ProjX in Bad block,
1328 * ProjX(Load) and ProjX(Store).
1330 static ir_node *equivalent_node_Proj(ir_node *proj) {
1331 ir_node *oldn = proj;
1332 ir_node *a = get_Proj_pred(proj);
1334 if (get_irn_op(a) == op_Tuple) {
1335 /* Remove the Tuple/Proj combination. */
1336 if ( get_Proj_proj(proj) <= get_Tuple_n_preds(a) ) {
1337 proj = get_Tuple_pred(a, get_Proj_proj(proj));
1338 DBG_OPT_TUPLE(oldn, a, proj);
1340 /* This should not happen! */
1341 assert(! "found a Proj with higher number than Tuple predecessors");
1344 } else if (get_irn_mode(proj) == mode_X) {
1345 if (is_Block_dead(get_nodes_block(skip_Proj(proj)))) {
1346 /* Remove dead control flow -- early gigo(). */
1348 } else if (get_opt_ldst_only_null_ptr_exceptions()) {
1349 ir_op *op = get_irn_op(a);
1351 if (op == op_Load) {
1352 /* get the Load address */
1353 ir_node *addr = get_Load_ptr(a);
1354 ir_node *blk = get_irn_n(a, -1);
1357 if (value_not_null(blk, addr, &confirm)) {
1358 if (confirm == NULL) {
1359 /* this node may float if it did not depend on a Confirm */
1360 set_irn_pinned(a, op_pin_state_floats);
1362 if (get_Proj_proj(proj) == pn_Load_X_except) {
1363 DBG_OPT_EXC_REM(proj);
1366 return new_r_Jmp(current_ir_graph, blk);
1368 } else if (op == op_Store) {
1369 /* get the load/store address */
1370 ir_node *addr = get_Store_ptr(a);
1371 ir_node *blk = get_irn_n(a, -1);
1374 if (value_not_null(blk, addr, &confirm)) {
1375 if (confirm == NULL) {
1376 /* this node may float if it did not depend on a Confirm */
1377 set_irn_pinned(a, op_pin_state_floats);
1379 if (get_Proj_proj(proj) == pn_Store_X_except) {
1380 DBG_OPT_EXC_REM(proj);
1383 return new_r_Jmp(current_ir_graph, blk);
1390 } /* equivalent_node_Proj */
1395 static ir_node *equivalent_node_Id(ir_node *n) {
1400 } while (get_irn_op(n) == op_Id);
1402 DBG_OPT_ID(oldn, n);
1404 } /* equivalent_node_Id */
1409 static ir_node *equivalent_node_Mux(ir_node *n)
1411 ir_node *oldn = n, *sel = get_Mux_sel(n);
1412 tarval *ts = value_of(sel);
1414 /* Mux(true, f, t) == t */
1415 if (ts == tarval_b_true) {
1416 n = get_Mux_true(n);
1417 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1419 /* Mux(false, f, t) == f */
1420 else if (ts == tarval_b_false) {
1421 n = get_Mux_false(n);
1422 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1424 /* Mux(v, x, x) == x */
1425 else if (get_Mux_false(n) == get_Mux_true(n)) {
1426 n = get_Mux_true(n);
1427 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1429 else if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(get_irn_mode(n))) {
1430 ir_node *cmp = get_Proj_pred(sel);
1431 long proj_nr = get_Proj_proj(sel);
1432 ir_node *b = get_Mux_false(n);
1433 ir_node *a = get_Mux_true(n);
1436 * Note: normalization puts the constant on the right site,
1437 * so we check only one case.
1439 * Note further that these optimization work even for floating point
1440 * with NaN's because -NaN == NaN.
1441 * However, if +0 and -0 is handled differently, we cannot use the first one.
1443 if (get_irn_op(cmp) == op_Cmp && get_Cmp_left(cmp) == a) {
1444 if (classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
1445 /* Mux(a CMP 0, X, a) */
1446 if (get_irn_op(b) == op_Minus && get_Minus_op(b) == a) {
1447 /* Mux(a CMP 0, -a, a) */
1448 if (proj_nr == pn_Cmp_Eq) {
1449 /* Mux(a == 0, -a, a) ==> -a */
1451 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1452 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1453 /* Mux(a != 0, -a, a) ==> a */
1455 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1457 } else if (classify_Const(b) == CNST_NULL) {
1458 /* Mux(a CMP 0, 0, a) */
1459 if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1460 /* Mux(a != 0, 0, a) ==> a */
1462 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1463 } else if (proj_nr == pn_Cmp_Eq) {
1464 /* Mux(a == 0, 0, a) ==> 0 */
1466 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1473 } /* equivalent_node_Mux */
1476 * Returns a equivalent node of a Psi: if a condition is true
1477 * and all previous conditions are false we know its value.
1478 * If all conditions are false its value is the default one.
1480 static ir_node *equivalent_node_Psi(ir_node *n) {
1482 return equivalent_node_Mux(n);
1484 } /* equivalent_node_Psi */
1487 * Optimize -a CMP -b into b CMP a.
1488 * This works only for for modes where unary Minus
1490 * Note that two-complement integers can Overflow
1491 * so it will NOT work.
1493 * For == and != can be handled in Proj(Cmp)
1495 static ir_node *equivalent_node_Cmp(ir_node *n) {
1496 ir_node *left = get_Cmp_left(n);
1497 ir_node *right = get_Cmp_right(n);
1499 if (get_irn_op(left) == op_Minus && get_irn_op(right) == op_Minus &&
1500 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1501 left = get_Minus_op(left);
1502 right = get_Minus_op(right);
1503 set_Cmp_left(n, right);
1504 set_Cmp_right(n, left);
1507 } /* equivalent_node_Cmp */
1510 * Remove Confirm nodes if setting is on.
1511 * Replace Confirms(x, '=', Constlike) by Constlike.
1513 static ir_node *equivalent_node_Confirm(ir_node *n) {
1514 ir_node *pred = get_Confirm_value(n);
1515 pn_Cmp pnc = get_Confirm_cmp(n);
1517 if (get_irn_op(pred) == op_Confirm && pnc == get_Confirm_cmp(pred)) {
1519 * rare case: two identical Confirms one after another,
1520 * replace the second one with the first.
1524 if (pnc == pn_Cmp_Eq) {
1525 ir_node *bound = get_Confirm_bound(n);
1528 * Optimize a rare case:
1529 * Confirm(x, '=', Constlike) ==> Constlike
1531 if (is_irn_constlike(bound)) {
1532 DBG_OPT_CONFIRM(n, bound);
1536 return get_opt_remove_confirm() ? get_Confirm_value(n) : n;
1540 * Optimize CopyB(mem, x, x) into a Nop.
1542 static ir_node *equivalent_node_CopyB(ir_node *n) {
1543 ir_node *a = get_CopyB_dst(n);
1544 ir_node *b = get_CopyB_src(n);
1547 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1548 ir_node *mem = get_CopyB_mem(n);
1549 ir_node *blk = get_nodes_block(n);
1550 turn_into_tuple(n, pn_CopyB_max);
1551 set_Tuple_pred(n, pn_CopyB_M, mem);
1552 set_Tuple_pred(n, pn_CopyB_X_regular, new_r_Jmp(current_ir_graph, blk));
1553 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1554 set_Tuple_pred(n, pn_CopyB_M_except, new_Bad());
1557 } /* equivalent_node_CopyB */
1560 * Optimize Bounds(idx, idx, upper) into idx.
1562 static ir_node *equivalent_node_Bound(ir_node *n) {
1563 ir_node *idx = get_Bound_index(n);
1564 ir_node *lower = get_Bound_lower(n);
1567 /* By definition lower < upper, so if idx == lower -->
1568 lower <= idx && idx < upper */
1570 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1573 ir_node *pred = skip_Proj(idx);
1575 if (get_irn_op(pred) == op_Bound) {
1577 * idx was Bounds_check previously, it is still valid if
1578 * lower <= pred_lower && pred_upper <= upper.
1580 ir_node *upper = get_Bound_upper(n);
1581 if (get_Bound_lower(pred) == lower &&
1582 get_Bound_upper(pred) == upper) {
1584 * One could expect that we simply return the previous
1585 * Bound here. However, this would be wrong, as we could
1586 * add an exception Proj to a new location than.
1587 * So, we must turn in into a tuple
1594 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1595 ir_node *mem = get_Bound_mem(n);
1596 ir_node *blk = get_nodes_block(n);
1597 turn_into_tuple(n, pn_Bound_max);
1598 set_Tuple_pred(n, pn_Bound_M, mem);
1599 set_Tuple_pred(n, pn_Bound_X_regular, new_r_Jmp(current_ir_graph, blk)); /* no exception */
1600 set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */
1601 set_Tuple_pred(n, pn_Bound_res, idx);
1604 } /* equivalent_node_Bound */
1607 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1608 * perform no actual computation, as, e.g., the Id nodes. It does not create
1609 * new nodes. It is therefore safe to free n if the node returned is not n.
1610 * If a node returns a Tuple we can not just skip it. If the size of the
1611 * in array fits, we transform n into a tuple (e.g., Div).
1613 ir_node *equivalent_node(ir_node *n) {
1614 if (n->op->ops.equivalent_node)
1615 return n->op->ops.equivalent_node(n);
1617 } /* equivalent_node */
1620 * Sets the default equivalent node operation for an ir_op_ops.
1622 * @param code the opcode for the default operation
1623 * @param ops the operations initialized
1628 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1632 ops->equivalent_node = equivalent_node_##a; \
1672 } /* firm_set_default_equivalent_node */
1675 * Do node specific optimizations of nodes predecessors.
1677 static void optimize_preds(ir_node *n) {
1678 switch (get_irn_opcode(n)) {
1680 case iro_Cmp: { /* We don't want Cast as input to Cmp. */
1681 ir_node *a = get_Cmp_left(n), *b = get_Cmp_right(n);
1683 if (get_irn_op(a) == op_Cast) {
1687 if (get_irn_op(b) == op_Cast) {
1689 set_Cmp_right(n, b);
1696 } /* optimize_preds */
1699 * Returns non-zero if a node is a Phi node
1700 * with all predecessors constant.
1702 static int is_const_Phi(ir_node *n) {
1707 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1708 if (! is_Const(get_irn_n(n, i)))
1711 } /* is_const_Phi */
1714 * Apply an evaluator on a binop with a constant operators (and one Phi).
1716 * @param phi the Phi node
1717 * @param other the other operand
1718 * @param eval an evaluator function
1719 * @param left if non-zero, other is the left operand, else the right
1721 * @return a new Phi node if the conversion was successful, NULL else
1723 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), int left) {
1729 int i, n = get_irn_arity(phi);
1731 NEW_ARR_A(void *, res, n);
1733 for (i = 0; i < n; ++i) {
1734 pred = get_irn_n(phi, i);
1735 tv = get_Const_tarval(pred);
1736 tv = eval(other, tv);
1738 if (tv == tarval_bad) {
1739 /* folding failed, bad */
1745 for (i = 0; i < n; ++i) {
1746 pred = get_irn_n(phi, i);
1747 tv = get_Const_tarval(pred);
1748 tv = eval(tv, other);
1750 if (tv == tarval_bad) {
1751 /* folding failed, bad */
1757 mode = get_irn_mode(phi);
1758 irg = current_ir_graph;
1759 for (i = 0; i < n; ++i) {
1760 pred = get_irn_n(phi, i);
1761 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1762 mode, res[i], get_Const_type(pred));
1764 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1765 } /* apply_binop_on_phi */
1768 * Apply an evaluator on a unop with a constant operator (a Phi).
1770 * @param phi the Phi node
1771 * @param eval an evaluator function
1773 * @return a new Phi node if the conversion was successful, NULL else
1775 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
1781 int i, n = get_irn_arity(phi);
1783 NEW_ARR_A(void *, res, n);
1784 for (i = 0; i < n; ++i) {
1785 pred = get_irn_n(phi, i);
1786 tv = get_Const_tarval(pred);
1789 if (tv == tarval_bad) {
1790 /* folding failed, bad */
1795 mode = get_irn_mode(phi);
1796 irg = current_ir_graph;
1797 for (i = 0; i < n; ++i) {
1798 pred = get_irn_n(phi, i);
1799 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1800 mode, res[i], get_Const_type(pred));
1802 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1803 } /* apply_unop_on_phi */
1806 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1807 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1808 * If possible, remove the Conv's.
1810 static ir_node *transform_node_AddSub(ir_node *n) {
1811 ir_mode *mode = get_irn_mode(n);
1813 if (mode_is_reference(mode)) {
1814 ir_node *left = get_binop_left(n);
1815 ir_node *right = get_binop_right(n);
1816 int ref_bits = get_mode_size_bits(mode);
1818 if (get_irn_op(left) == op_Conv) {
1819 ir_mode *mode = get_irn_mode(left);
1820 int bits = get_mode_size_bits(mode);
1822 if (ref_bits == bits &&
1823 mode_is_int(mode) &&
1824 get_mode_arithmetic(mode) == irma_twos_complement) {
1825 ir_node *pre = get_Conv_op(left);
1826 ir_mode *pre_mode = get_irn_mode(pre);
1828 if (mode_is_int(pre_mode) &&
1829 get_mode_size_bits(pre_mode) == bits &&
1830 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1831 /* ok, this conv just changes to sign, moreover the calculation
1832 * is done with same number of bits as our address mode, so
1833 * we can ignore the conv as address calculation can be viewed
1834 * as either signed or unsigned
1836 set_binop_left(n, pre);
1841 if (get_irn_op(right) == op_Conv) {
1842 ir_mode *mode = get_irn_mode(right);
1843 int bits = get_mode_size_bits(mode);
1845 if (ref_bits == bits &&
1846 mode_is_int(mode) &&
1847 get_mode_arithmetic(mode) == irma_twos_complement) {
1848 ir_node *pre = get_Conv_op(right);
1849 ir_mode *pre_mode = get_irn_mode(pre);
1851 if (mode_is_int(pre_mode) &&
1852 get_mode_size_bits(pre_mode) == bits &&
1853 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1854 /* ok, this conv just changes to sign, moreover the calculation
1855 * is done with same number of bits as our address mode, so
1856 * we can ignore the conv as address calculation can be viewed
1857 * as either signed or unsigned
1859 set_binop_right(n, pre);
1865 } /* transform_node_AddSub */
1867 #define HANDLE_BINOP_PHI(op,a,b,c) \
1869 if (is_Const(b) && is_const_Phi(a)) { \
1870 /* check for Op(Phi, Const) */ \
1871 c = apply_binop_on_phi(a, get_Const_tarval(b), op, 0); \
1873 else if (is_Const(a) && is_const_Phi(b)) { \
1874 /* check for Op(Const, Phi) */ \
1875 c = apply_binop_on_phi(b, get_Const_tarval(a), op, 1); \
1878 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1882 #define HANDLE_UNOP_PHI(op,a,c) \
1884 if (is_const_Phi(a)) { \
1885 /* check for Op(Phi) */ \
1886 c = apply_unop_on_phi(a, op); \
1889 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1895 * Do the AddSub optimization, then Transform
1896 * Constant folding on Phi
1897 * Add(a,a) -> Mul(a, 2)
1898 * Add(Mul(a, x), a) -> Mul(a, x+1)
1899 * if the mode is integer or float.
1900 * Transform Add(a,-b) into Sub(a,b).
1901 * Reassociation might fold this further.
1903 static ir_node *transform_node_Add(ir_node *n) {
1905 ir_node *a, *b, *c, *oldn = n;
1907 n = transform_node_AddSub(n);
1909 a = get_Add_left(n);
1910 b = get_Add_right(n);
1912 HANDLE_BINOP_PHI(tarval_add, a,b,c);
1914 mode = get_irn_mode(n);
1916 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1917 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1920 if (mode_is_num(mode)) {
1922 ir_node *block = get_irn_n(n, -1);
1925 get_irn_dbg_info(n),
1929 new_r_Const_long(current_ir_graph, block, mode, 2),
1931 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
1932 } else if (get_irn_op(a) == op_Minus) {
1934 get_irn_dbg_info(n),
1940 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1941 } else if (get_irn_op(b) == op_Minus) {
1943 get_irn_dbg_info(n),
1949 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1951 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1952 else if (!get_opt_reassociation() && get_irn_op(a) == op_Mul) {
1953 ir_node *ma = get_Mul_left(a);
1954 ir_node *mb = get_Mul_right(a);
1957 ir_node *blk = get_irn_n(n, -1);
1959 get_irn_dbg_info(n), current_ir_graph, blk,
1962 get_irn_dbg_info(n), current_ir_graph, blk,
1964 new_r_Const_long(current_ir_graph, blk, mode, 1),
1967 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1968 } else if (b == mb) {
1969 ir_node *blk = get_irn_n(n, -1);
1971 get_irn_dbg_info(n), current_ir_graph, blk,
1974 get_irn_dbg_info(n), current_ir_graph, blk,
1976 new_r_Const_long(current_ir_graph, blk, mode, 1),
1979 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1982 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1983 else if (!get_opt_reassociation() && get_irn_op(b) == op_Mul) {
1984 ir_node *ma = get_Mul_left(b);
1985 ir_node *mb = get_Mul_right(b);
1988 ir_node *blk = get_irn_n(n, -1);
1990 get_irn_dbg_info(n), current_ir_graph, blk,
1993 get_irn_dbg_info(n), current_ir_graph, blk,
1995 new_r_Const_long(current_ir_graph, blk, mode, 1),
1998 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1999 } else if (a == mb) {
2000 ir_node *blk = get_irn_n(n, -1);
2002 get_irn_dbg_info(n), current_ir_graph, blk,
2005 get_irn_dbg_info(n), current_ir_graph, blk,
2007 new_r_Const_long(current_ir_graph, blk, mode, 1),
2010 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
2015 } /* transform_node_Add */
2018 * Do the AddSub optimization, then Transform
2019 * Constant folding on Phi
2020 * Sub(0,a) -> Minus(a)
2021 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2022 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2024 static ir_node *transform_node_Sub(ir_node *n) {
2029 n = transform_node_AddSub(n);
2031 a = get_Sub_left(n);
2032 b = get_Sub_right(n);
2034 HANDLE_BINOP_PHI(tarval_sub, a,b,c);
2036 mode = get_irn_mode(n);
2038 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2039 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2042 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2043 if (mode_is_num(mode) && mode == get_irn_mode(a) && (classify_Const(a) == CNST_NULL)) {
2045 get_irn_dbg_info(n),
2050 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2052 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2053 else if (get_opt_reassociation() && get_irn_op(a) == op_Mul) {
2054 ir_node *ma = get_Mul_left(a);
2055 ir_node *mb = get_Mul_right(a);
2058 ir_node *blk = get_irn_n(n, -1);
2060 get_irn_dbg_info(n),
2061 current_ir_graph, blk,
2064 get_irn_dbg_info(n),
2065 current_ir_graph, blk,
2067 new_r_Const_long(current_ir_graph, blk, mode, 1),
2070 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2071 } else if (mb == b) {
2072 ir_node *blk = get_irn_n(n, -1);
2074 get_irn_dbg_info(n),
2075 current_ir_graph, blk,
2078 get_irn_dbg_info(n),
2079 current_ir_graph, blk,
2081 new_r_Const_long(current_ir_graph, blk, mode, 1),
2084 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2086 } else if (get_irn_op(a) == op_Sub) {
2087 ir_node *x = get_Sub_left(a);
2088 ir_node *y = get_Sub_right(a);
2089 ir_node *blk = get_irn_n(n, -1);
2090 ir_mode *m_b = get_irn_mode(b);
2091 ir_mode *m_y = get_irn_mode(y);
2094 /* Determine the right mode for the Add. */
2097 else if (mode_is_reference(m_b))
2099 else if (mode_is_reference(m_y))
2103 * Both modes are different but none is reference,
2104 * happens for instance in SubP(SubP(P, Iu), Is).
2105 * We have two possibilities here: Cast or ignore.
2106 * Currently we ignore this case.
2111 add = new_r_Add(current_ir_graph, blk, y, b, mode);
2114 set_Sub_right(n, add);
2115 DBG_OPT_ALGSIM0(n, n, FS_OPT_SUB_SUB_X_Y_Z);
2119 } /* transform_node_Sub */
2122 * Transform Mul(a,-1) into -a.
2123 * Do constant evaluation of Phi nodes.
2124 * Do architecture dependent optimizations on Mul nodes
2126 static ir_node *transform_node_Mul(ir_node *n) {
2127 ir_node *c, *oldn = n;
2128 ir_node *a = get_Mul_left(n);
2129 ir_node *b = get_Mul_right(n);
2132 HANDLE_BINOP_PHI(tarval_mul, a,b,c);
2134 mode = get_irn_mode(n);
2135 if (mode_is_signed(mode)) {
2138 if (value_of(a) == get_mode_minus_one(mode))
2140 else if (value_of(b) == get_mode_minus_one(mode))
2143 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2144 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2148 return arch_dep_replace_mul_with_shifts(n);
2149 } /* transform_node_Mul */
2152 * Transform a Div Node.
2154 static ir_node *transform_node_Div(ir_node *n) {
2155 tarval *tv = value_of(n);
2158 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2160 if (tv != tarval_bad) {
2161 value = new_Const(get_tarval_mode(tv), tv);
2163 DBG_OPT_CSTEVAL(n, value);
2164 } else /* Try architecture dependent optimization */
2165 value = arch_dep_replace_div_by_const(n);
2168 /* Turn Div into a tuple (mem, jmp, bad, value) */
2169 ir_node *mem = get_Div_mem(n);
2170 ir_node *blk = get_nodes_block(n);
2172 turn_into_tuple(n, pn_Div_max);
2173 set_Tuple_pred(n, pn_Div_M, mem);
2174 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2175 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2176 set_Tuple_pred(n, pn_Div_res, value);
2179 } /* transform_node_Div */
2182 * Transform a Mod node.
2184 static ir_node *transform_node_Mod(ir_node *n) {
2185 tarval *tv = value_of(n);
2188 /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */
2190 if (tv != tarval_bad) {
2191 value = new_Const(get_tarval_mode(tv), tv);
2193 DBG_OPT_CSTEVAL(n, value);
2194 } else /* Try architecture dependent optimization */
2195 value = arch_dep_replace_mod_by_const(n);
2198 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2199 ir_node *mem = get_Mod_mem(n);
2200 ir_node *blk = get_nodes_block(n);
2202 turn_into_tuple(n, pn_Mod_max);
2203 set_Tuple_pred(n, pn_Mod_M, mem);
2204 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2205 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2206 set_Tuple_pred(n, pn_Mod_res, value);
2209 } /* transform_node_Mod */
2212 * Transform a DivMod node.
2214 static ir_node *transform_node_DivMod(ir_node *n) {
2217 ir_node *a = get_DivMod_left(n);
2218 ir_node *b = get_DivMod_right(n);
2219 ir_mode *mode = get_irn_mode(a);
2220 tarval *ta = value_of(a);
2221 tarval *tb = value_of(b);
2223 if (!(mode_is_int(mode) && mode_is_int(get_irn_mode(b))))
2226 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2228 if (tb != tarval_bad) {
2229 if (tb == get_mode_one(get_tarval_mode(tb))) {
2230 b = new_Const (mode, get_mode_null(mode));
2233 DBG_OPT_CSTEVAL(n, b);
2234 } else if (ta != tarval_bad) {
2235 tarval *resa, *resb;
2236 resa = tarval_div (ta, tb);
2237 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2238 Jmp for X result!? */
2239 resb = tarval_mod (ta, tb);
2240 if (resb == tarval_bad) return n; /* Causes exception! */
2241 a = new_Const (mode, resa);
2242 b = new_Const (mode, resb);
2245 DBG_OPT_CSTEVAL(n, a);
2246 DBG_OPT_CSTEVAL(n, b);
2247 } else { /* Try architecture dependent optimization */
2248 arch_dep_replace_divmod_by_const(&a, &b, n);
2249 evaluated = a != NULL;
2251 } else if (ta == get_mode_null(mode)) {
2252 /* 0 / non-Const = 0 */
2257 if (evaluated) { /* replace by tuple */
2258 ir_node *mem = get_DivMod_mem(n);
2259 ir_node *blk = get_nodes_block(n);
2260 turn_into_tuple(n, pn_DivMod_max);
2261 set_Tuple_pred(n, pn_DivMod_M, mem);
2262 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
2263 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
2264 set_Tuple_pred(n, pn_DivMod_res_div, a);
2265 set_Tuple_pred(n, pn_DivMod_res_mod, b);
2269 } /* transform_node_DivMod */
2272 * Optimize Abs(x) into x if x is Confirmed >= 0
2273 * Optimize Abs(x) into -x if x is Confirmed <= 0
2275 static ir_node *transform_node_Abs(ir_node *n) {
2277 ir_node *a = get_Abs_op(n);
2278 value_classify_sign sign = classify_value_sign(a);
2280 if (sign == value_classified_negative) {
2281 ir_mode *mode = get_irn_mode(n);
2284 * We can replace the Abs by -x here.
2285 * We even could add a new Confirm here.
2287 * Note that -x would create a new node, so we could
2288 * not run it in the equivalent_node() context.
2290 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
2291 get_irn_n(n, -1), a, mode);
2293 DBG_OPT_CONFIRM(oldn, n);
2294 } else if (sign == value_classified_positive) {
2295 /* n is positive, Abs is not needed */
2298 DBG_OPT_CONFIRM(oldn, n);
2302 } /* transform_node_Abs */
2305 * Transform a Cond node.
2307 * Replace the Cond by a Jmp if it branches on a constant
2310 static ir_node *transform_node_Cond(ir_node *n) {
2313 ir_node *a = get_Cond_selector(n);
2314 tarval *ta = value_of(a);
2316 /* we need block info which is not available in floating irgs */
2317 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
2320 if ((ta != tarval_bad) &&
2321 (get_irn_mode(a) == mode_b) &&
2322 (get_opt_unreachable_code())) {
2323 /* It's a boolean Cond, branching on a boolean constant.
2324 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2325 jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n));
2326 turn_into_tuple(n, pn_Cond_max);
2327 if (ta == tarval_b_true) {
2328 set_Tuple_pred(n, pn_Cond_false, new_Bad());
2329 set_Tuple_pred(n, pn_Cond_true, jmp);
2331 set_Tuple_pred(n, pn_Cond_false, jmp);
2332 set_Tuple_pred(n, pn_Cond_true, new_Bad());
2334 /* We might generate an endless loop, so keep it alive. */
2335 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
2338 } /* transform_node_Cond */
2343 static ir_node *transform_node_And(ir_node *n) {
2344 ir_node *c, *oldn = n;
2345 ir_node *a = get_And_left(n);
2346 ir_node *b = get_And_right(n);
2348 HANDLE_BINOP_PHI(tarval_and, a,b,c);
2350 } /* transform_node_And */
2355 static ir_node *transform_node_Eor(ir_node *n) {
2356 ir_node *c, *oldn = n;
2357 ir_node *a = get_Eor_left(n);
2358 ir_node *b = get_Eor_right(n);
2359 ir_mode *mode = get_irn_mode(n);
2361 HANDLE_BINOP_PHI(tarval_eor, a,b,c);
2365 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
2366 mode, get_mode_null(mode));
2367 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
2368 } else if ((mode == mode_b)
2369 && (get_irn_op(a) == op_Proj)
2370 && (get_irn_mode(a) == mode_b)
2371 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)
2372 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2373 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
2374 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2375 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
2377 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
2378 } else if ((mode == mode_b)
2379 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) {
2380 /* The Eor is a Not. Replace it by a Not. */
2381 /* ????!!!Extend to bitfield 1111111. */
2382 n = new_r_Not(current_ir_graph, get_irn_n(n, -1), a, mode_b);
2384 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2388 } /* transform_node_Eor */
2393 static ir_node *transform_node_Not(ir_node *n) {
2394 ir_node *c, *oldn = n;
2395 ir_node *a = get_Not_op(n);
2397 HANDLE_UNOP_PHI(tarval_not,a,c);
2399 /* check for a boolean Not */
2400 if ( (get_irn_mode(n) == mode_b)
2401 && (get_irn_op(a) == op_Proj)
2402 && (get_irn_mode(a) == mode_b)
2403 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2404 /* We negate a Cmp. The Cmp has the negated result anyways! */
2405 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2406 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
2407 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
2410 } /* transform_node_Not */
2413 * Transform a Minus.
2415 static ir_node *transform_node_Minus(ir_node *n) {
2416 ir_node *c, *oldn = n;
2417 ir_node *a = get_Minus_op(n);
2419 HANDLE_UNOP_PHI(tarval_neg,a,c);
2421 } /* transform_node_Minus */
2424 * Transform a Cast_type(Const) into a new Const_type
2426 static ir_node *transform_node_Cast(ir_node *n) {
2428 ir_node *pred = get_Cast_op(n);
2429 ir_type *tp = get_irn_type(n);
2431 if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) {
2432 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
2433 get_Const_tarval(pred), tp);
2434 DBG_OPT_CSTEVAL(oldn, n);
2435 } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) {
2436 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred),
2437 get_SymConst_kind(pred), tp);
2438 DBG_OPT_CSTEVAL(oldn, n);
2442 } /* transform_node_Cast */
2445 * Transform a Proj(Div) with a non-zero value.
2446 * Removes the exceptions and routes the memory to the NoMem node.
2448 static ir_node *transform_node_Proj_Div(ir_node *proj) {
2449 ir_node *div = get_Proj_pred(proj);
2450 ir_node *blk = get_nodes_block(div);
2451 ir_node *b = get_Div_right(div);
2455 if (value_not_zero(blk, b, &confirm)) {
2456 /* div(x, y) && y != 0 */
2457 proj_nr = get_Proj_proj(proj);
2458 if (proj_nr == pn_Div_X_except) {
2459 /* we found an exception handler, remove it */
2460 DBG_OPT_EXC_REM(proj);
2462 } else if (proj_nr == pn_Div_M) {
2463 ir_node *res = get_Div_mem(div);
2464 ir_node *new_mem = get_irg_no_mem(current_ir_graph);
2467 /* This node can only float up to the Confirm block */
2468 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2470 set_irn_pinned(div, op_pin_state_floats);
2471 /* this is a Div without exception, we can remove the memory edge */
2472 set_Div_mem(div, new_mem);
2477 } /* transform_node_Proj_Div */
2480 * Transform a Proj(Mod) with a non-zero value.
2481 * Removes the exceptions and routes the memory to the NoMem node.
2483 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
2484 ir_node *mod = get_Proj_pred(proj);
2485 ir_node *blk = get_nodes_block(mod);
2486 ir_node *b = get_Mod_right(mod);
2490 if (value_not_zero(blk, b, &confirm)) {
2491 /* mod(x, y) && y != 0 */
2492 proj_nr = get_Proj_proj(proj);
2494 if (proj_nr == pn_Mod_X_except) {
2495 /* we found an exception handler, remove it */
2496 DBG_OPT_EXC_REM(proj);
2498 } else if (proj_nr == pn_Mod_M) {
2499 ir_node *res = get_Mod_mem(mod);
2500 ir_node *new_mem = get_irg_no_mem(current_ir_graph);
2503 /* This node can only float up to the Confirm block */
2504 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2506 set_irn_pinned(mod, op_pin_state_floats);
2507 /* this is a Mod without exception, we can remove the memory edge */
2508 set_Mod_mem(mod, get_irg_no_mem(current_ir_graph));
2510 } else if (proj_nr == pn_Mod_res && get_Mod_left(mod) == b) {
2511 /* a % a = 0 if a != 0 */
2512 ir_mode *mode = get_irn_mode(proj);
2513 ir_node *res = new_Const(mode, get_mode_null(mode));
2515 DBG_OPT_CSTEVAL(mod, res);
2520 } /* transform_node_Proj_Mod */
2523 * Transform a Proj(DivMod) with a non-zero value.
2524 * Removes the exceptions and routes the memory to the NoMem node.
2526 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
2527 ir_node *divmod = get_Proj_pred(proj);
2528 ir_node *blk = get_nodes_block(divmod);
2529 ir_node *b = get_DivMod_right(divmod);
2533 if (value_not_zero(blk, b, &confirm)) {
2534 /* DivMod(x, y) && y != 0 */
2535 proj_nr = get_Proj_proj(proj);
2537 if (proj_nr == pn_DivMod_X_except) {
2538 /* we found an exception handler, remove it */
2539 DBG_OPT_EXC_REM(proj);
2541 } else if (proj_nr == pn_DivMod_M) {
2542 ir_node *res = get_DivMod_mem(divmod);
2543 ir_node *new_mem = get_irg_no_mem(current_ir_graph);
2546 /* This node can only float up to the Confirm block */
2547 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2549 set_irn_pinned(divmod, op_pin_state_floats);
2550 /* this is a DivMod without exception, we can remove the memory edge */
2551 set_DivMod_mem(divmod, get_irg_no_mem(current_ir_graph));
2553 } else if (proj_nr == pn_DivMod_res_mod && get_DivMod_left(divmod) == b) {
2554 /* a % a = 0 if a != 0 */
2555 ir_mode *mode = get_irn_mode(proj);
2556 ir_node *res = new_Const(mode, get_mode_null(mode));
2558 DBG_OPT_CSTEVAL(divmod, res);
2563 } /* transform_node_Proj_DivMod */
2566 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
2568 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
2569 if (get_opt_unreachable_code()) {
2570 ir_node *n = get_Proj_pred(proj);
2571 ir_node *b = get_Cond_selector(n);
2573 if (mode_is_int(get_irn_mode(b))) {
2574 tarval *tb = value_of(b);
2576 if (tb != tarval_bad) {
2577 /* we have a constant switch */
2578 long num = get_Proj_proj(proj);
2580 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
2581 if (get_tarval_long(tb) == num) {
2582 /* Do NOT create a jump here, or we will have 2 control flow ops
2583 * in a block. This case is optimized away in optimize_cf(). */
2586 /* this case will NEVER be taken, kill it */
2594 } /* transform_node_Proj_Cond */
2597 * Normalizes and optimizes Cmp nodes.
2599 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
2600 if (get_opt_reassociation()) {
2601 ir_node *n = get_Proj_pred(proj);
2602 ir_node *left = get_Cmp_left(n);
2603 ir_node *right = get_Cmp_right(n);
2607 ir_mode *mode = NULL;
2608 long proj_nr = get_Proj_proj(proj);
2611 * First step: normalize the compare op
2612 * by placing the constant on the right site
2613 * or moving the lower address node to the left.
2614 * We ignore the case that both are constants
2615 * this case should be optimized away.
2617 if (get_irn_op(right) == op_Const) {
2619 } else if (get_irn_op(left) == op_Const) {
2624 proj_nr = get_inversed_pnc(proj_nr);
2626 } else if (get_irn_idx(left) > get_irn_idx(right)) {
2632 proj_nr = get_inversed_pnc(proj_nr);
2637 * Second step: Try to reduce the magnitude
2638 * of a constant. This may help to generate better code
2639 * later and may help to normalize more compares.
2640 * Of course this is only possible for integer values.
2643 mode = get_irn_mode(c);
2644 tv = get_Const_tarval(c);
2646 if (tv != tarval_bad) {
2647 /* the following optimization is possible on modes without Overflow
2648 * on Unary Minus or on == and !=:
2649 * -a CMP c ==> a swap(CMP) -c
2651 * Beware: for two-complement Overflow may occur, so only == and != can
2652 * be optimized, see this:
2653 * -MININT < 0 =/=> MININT > 0 !!!
2655 if (get_opt_constant_folding() && get_irn_op(left) == op_Minus &&
2656 (!mode_overflow_on_unary_Minus(mode) ||
2657 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
2658 left = get_Minus_op(left);
2659 tv = tarval_sub(get_mode_null(mode), tv);
2661 proj_nr = get_inversed_pnc(proj_nr);
2665 /* for integer modes, we have more */
2666 if (mode_is_int(mode)) {
2667 /* Ne includes Unordered which is not possible on integers.
2668 * However, frontends often use this wrong, so fix it here */
2669 if (proj_nr & pn_Cmp_Uo) {
2670 proj_nr &= ~pn_Cmp_Uo;
2671 set_Proj_proj(proj, proj_nr);
2674 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
2675 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
2676 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
2677 tv = tarval_sub(tv, get_mode_one(mode));
2679 proj_nr ^= pn_Cmp_Eq;
2682 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
2683 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
2684 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
2685 tv = tarval_add(tv, get_mode_one(mode));
2687 proj_nr ^= pn_Cmp_Eq;
2691 /* the following reassociations work only for == and != */
2692 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
2694 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
2695 if (classify_tarval(tv) == TV_CLASSIFY_NULL && get_irn_op(left) == op_Sub) {
2696 right = get_Sub_right(left);
2697 left = get_Sub_left(left);
2699 tv = value_of(right);
2703 if (tv != tarval_bad) {
2704 ir_op *op = get_irn_op(left);
2706 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
2708 ir_node *c1 = get_Sub_right(left);
2709 tarval *tv2 = value_of(c1);
2711 if (tv2 != tarval_bad) {
2712 tv2 = tarval_add(tv, value_of(c1));
2714 if (tv2 != tarval_bad) {
2715 left = get_Sub_left(left);
2721 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
2722 else if (op == op_Add) {
2723 ir_node *a_l = get_Add_left(left);
2724 ir_node *a_r = get_Add_right(left);
2728 if (get_irn_op(a_l) == op_Const) {
2730 tv2 = value_of(a_l);
2733 tv2 = value_of(a_r);
2736 if (tv2 != tarval_bad) {
2737 tv2 = tarval_sub(tv, tv2);
2739 if (tv2 != tarval_bad) {
2746 /* -a == c ==> a == -c, -a != c ==> a != -c */
2747 else if (op == op_Minus) {
2748 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
2750 if (tv2 != tarval_bad) {
2751 left = get_Minus_op(left);
2758 /* the following reassociations work only for <= */
2759 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2760 if (tv != tarval_bad) {
2761 ir_op *op = get_irn_op(left);
2763 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
2771 * optimization for AND:
2773 * And(x, C) == C ==> And(x, C) != 0
2774 * And(x, C) != C ==> And(X, C) == 0
2776 * if C is a single Bit constant.
2778 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) &&
2779 (get_irn_op(left) == op_And)) {
2780 if (is_single_bit_tarval(tv)) {
2781 /* check for Constant's match. We have check hare the tarvals,
2782 because our const might be changed */
2783 ir_node *la = get_And_left(left);
2784 ir_node *ra = get_And_right(left);
2785 if ((is_Const(la) && get_Const_tarval(la) == tv) ||
2786 (is_Const(ra) && get_Const_tarval(ra) == tv)) {
2787 /* fine: do the transformation */
2788 tv = get_mode_null(get_tarval_mode(tv));
2789 proj_nr ^= pn_Cmp_Leg;
2794 } /* tarval != bad */
2798 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
2800 if (changed & 2) /* need a new Const */
2801 right = new_Const(mode, tv);
2803 /* create a new compare */
2804 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block,
2807 set_Proj_pred(proj, n);
2808 set_Proj_proj(proj, proj_nr);
2812 } /* transform_node_Proj_Cmp */
2815 * Does all optimizations on nodes that must be done on it's Proj's
2816 * because of creating new nodes.
2818 static ir_node *transform_node_Proj(ir_node *proj) {
2819 ir_node *n = get_Proj_pred(proj);
2821 switch (get_irn_opcode(n)) {
2823 return transform_node_Proj_Div(proj);
2826 return transform_node_Proj_Mod(proj);
2829 return transform_node_Proj_DivMod(proj);
2832 return transform_node_Proj_Cond(proj);
2835 return transform_node_Proj_Cmp(proj);
2838 /* should not happen, but if it does will be optimized away */
2839 return equivalent_node_Proj(proj);
2845 } /* transform_node_Proj */
2848 * Move Confirms down through Phi nodes.
2850 static ir_node *transform_node_Phi(ir_node *phi) {
2852 ir_mode *mode = get_irn_mode(phi);
2854 if (mode_is_reference(mode)) {
2855 n = get_irn_arity(phi);
2857 /* Beware of Phi0 */
2859 ir_node *pred = get_irn_n(phi, 0);
2860 ir_node *bound, *new_Phi, *block, **in;
2863 if (! is_Confirm(pred))
2866 bound = get_Confirm_bound(pred);
2867 pnc = get_Confirm_cmp(pred);
2869 NEW_ARR_A(ir_node *, in, n);
2870 in[0] = get_Confirm_value(pred);
2872 for (i = 1; i < n; ++i) {
2873 pred = get_irn_n(phi, i);
2875 if (! is_Confirm(pred) ||
2876 get_Confirm_bound(pred) != bound ||
2877 get_Confirm_cmp(pred) != pnc)
2879 in[i] = get_Confirm_value(pred);
2881 /* move the Confirm nodes "behind" the Phi */
2882 block = get_irn_n(phi, -1);
2883 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
2884 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
2888 } /* transform_node_Phi */
2891 * Returns the operands of a commutative bin-op, if one operand is
2892 * a const, it is returned as the second one.
2894 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
2895 ir_node *op_a = get_binop_left(binop);
2896 ir_node *op_b = get_binop_right(binop);
2898 assert(is_op_commutative(get_irn_op(binop)));
2900 if (get_irn_op(op_a) == op_Const) {
2907 } /* get_comm_Binop_Ops */
2910 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
2911 * Such pattern may arise in bitfield stores.
2913 * value c4 value c4 & c2
2914 * AND c3 AND c1 | c3
2919 static ir_node *transform_node_Or_bf_store(ir_node *or) {
2922 ir_node *and_l, *c3;
2923 ir_node *value, *c4;
2924 ir_node *new_and, *new_const, *block;
2925 ir_mode *mode = get_irn_mode(or);
2927 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
2929 get_comm_Binop_Ops(or, &and, &c1);
2930 if ((get_irn_op(c1) != op_Const) || (get_irn_op(and) != op_And))
2933 get_comm_Binop_Ops(and, &or_l, &c2);
2934 if ((get_irn_op(c2) != op_Const) || (get_irn_op(or_l) != op_Or))
2937 get_comm_Binop_Ops(or_l, &and_l, &c3);
2938 if ((get_irn_op(c3) != op_Const) || (get_irn_op(and_l) != op_And))
2941 get_comm_Binop_Ops(and_l, &value, &c4);
2942 if (get_irn_op(c4) != op_Const)
2945 /* ok, found the pattern, check for conditions */
2946 assert(mode == get_irn_mode(and));
2947 assert(mode == get_irn_mode(or_l));
2948 assert(mode == get_irn_mode(and_l));
2950 tv1 = get_Const_tarval(c1);
2951 tv2 = get_Const_tarval(c2);
2952 tv3 = get_Const_tarval(c3);
2953 tv4 = get_Const_tarval(c4);
2955 tv = tarval_or(tv4, tv2);
2956 if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) {
2957 /* have at least one 0 at the same bit position */
2961 n_tv4 = tarval_not(tv4);
2962 if (tv3 != tarval_and(tv3, n_tv4)) {
2963 /* bit in the or_mask is outside the and_mask */
2967 n_tv2 = tarval_not(tv2);
2968 if (tv1 != tarval_and(tv1, n_tv2)) {
2969 /* bit in the or_mask is outside the and_mask */
2973 /* ok, all conditions met */
2974 block = get_irn_n(or, -1);
2976 new_and = new_r_And(current_ir_graph, block,
2977 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
2979 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
2981 set_Or_left(or, new_and);
2982 set_Or_right(or, new_const);
2984 /* check for more */
2985 return transform_node_Or_bf_store(or);
2986 } /* transform_node_Or_bf_store */
2989 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
2991 static ir_node *transform_node_Or_Rot(ir_node *or) {
2992 ir_mode *mode = get_irn_mode(or);
2993 ir_node *shl, *shr, *block;
2994 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
2997 if (! mode_is_int(mode))
3000 shl = get_binop_left(or);
3001 shr = get_binop_right(or);
3003 if (get_irn_op(shl) == op_Shr) {
3004 if (get_irn_op(shr) != op_Shl)
3010 } else if (get_irn_op(shl) != op_Shl) {
3012 } else if (get_irn_op(shr) != op_Shr) {
3015 x = get_Shl_left(shl);
3016 if (x != get_Shr_left(shr))
3019 c1 = get_Shl_right(shl);
3020 c2 = get_Shr_right(shr);
3021 if (get_irn_op(c1) == op_Const && get_irn_op(c2) == op_Const) {
3022 tv1 = get_Const_tarval(c1);
3023 if (! tarval_is_long(tv1))
3026 tv2 = get_Const_tarval(c2);
3027 if (! tarval_is_long(tv2))
3030 if (get_tarval_long(tv1) + get_tarval_long(tv2)
3031 != get_mode_size_bits(mode))
3034 /* yet, condition met */
3035 block = get_irn_n(or, -1);
3037 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
3039 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
3041 } else if (get_irn_op(c1) == op_Sub) {
3045 if (get_Sub_right(sub) != v)
3048 c1 = get_Sub_left(sub);
3049 if (get_irn_op(c1) != op_Const)
3052 tv1 = get_Const_tarval(c1);
3053 if (! tarval_is_long(tv1))
3056 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
3059 /* yet, condition met */
3060 block = get_nodes_block(or);
3062 /* a Rot right is not supported, so use a rot left */
3063 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
3065 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
3067 } else if (get_irn_op(c2) == op_Sub) {
3071 c1 = get_Sub_left(sub);
3072 if (get_irn_op(c1) != op_Const)
3075 tv1 = get_Const_tarval(c1);
3076 if (! tarval_is_long(tv1))
3079 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
3082 /* yet, condition met */
3083 block = get_irn_n(or, -1);
3086 n = new_r_Rot(current_ir_graph, block, x, v, mode);
3088 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
3093 } /* transform_node_Or_Rot */
3098 static ir_node *transform_node_Or(ir_node *n) {
3099 ir_node *c, *oldn = n;
3100 ir_node *a = get_Or_left(n);
3101 ir_node *b = get_Or_right(n);
3103 HANDLE_BINOP_PHI(tarval_or, a,b,c);
3105 n = transform_node_Or_bf_store(n);
3106 n = transform_node_Or_Rot(n);
3109 } /* transform_node_Or */
3113 static ir_node *transform_node(ir_node *n);
3116 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl.
3118 * Should be moved to reassociation?
3120 static ir_node *transform_node_shift(ir_node *n) {
3121 ir_node *left, *right;
3122 tarval *tv1, *tv2, *res;
3124 int modulo_shf, flag;
3126 left = get_binop_left(n);
3128 /* different operations */
3129 if (get_irn_op(left) != get_irn_op(n))
3132 right = get_binop_right(n);
3133 tv1 = value_of(right);
3134 if (tv1 == tarval_bad)
3137 tv2 = value_of(get_binop_right(left));
3138 if (tv2 == tarval_bad)
3141 res = tarval_add(tv1, tv2);
3143 /* beware: a simple replacement works only, if res < modulo shift */
3144 mode = get_irn_mode(n);
3148 modulo_shf = get_mode_modulo_shift(mode);
3149 if (modulo_shf > 0) {
3150 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
3152 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
3158 /* ok, we can replace it */
3159 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
3161 in[0] = get_binop_left(left);
3162 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
3164 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
3166 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
3168 return transform_node(irn);
3171 } /* transform_node_shift */
3176 static ir_node *transform_node_Shr(ir_node *n) {
3177 ir_node *c, *oldn = n;
3178 ir_node *a = get_Shr_left(n);
3179 ir_node *b = get_Shr_right(n);
3181 HANDLE_BINOP_PHI(tarval_shr, a, b, c);
3182 return transform_node_shift(n);
3183 } /* transform_node_Shr */
3188 static ir_node *transform_node_Shrs(ir_node *n) {
3189 ir_node *c, *oldn = n;
3190 ir_node *a = get_Shrs_left(n);
3191 ir_node *b = get_Shrs_right(n);
3193 HANDLE_BINOP_PHI(tarval_shrs, a, b, c);
3194 return transform_node_shift(n);
3195 } /* transform_node_Shrs */
3200 static ir_node *transform_node_Shl(ir_node *n) {
3201 ir_node *c, *oldn = n;
3202 ir_node *a = get_Shl_left(n);
3203 ir_node *b = get_Shl_right(n);
3205 HANDLE_BINOP_PHI(tarval_shl, a, b, c);
3206 return transform_node_shift(n);
3207 } /* transform_node_Shl */
3210 * Remove dead blocks and nodes in dead blocks
3211 * in keep alive list. We do not generate a new End node.
3213 static ir_node *transform_node_End(ir_node *n) {
3214 int i, j, n_keepalives = get_End_n_keepalives(n);
3217 NEW_ARR_A(ir_node *, in, n_keepalives);
3219 for (i = j = 0; i < n_keepalives; ++i) {
3220 ir_node *ka = get_End_keepalive(n, i);
3222 if (! is_Block_dead(ka)) {
3226 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
3229 /* FIXME: beabi need to keep a Proj(M) */
3230 if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka))
3233 if (j != n_keepalives)
3234 set_End_keepalives(n, j, in);
3236 } /* transform_node_End */
3239 * Optimize a Mux into some simpler cases.
3241 static ir_node *transform_node_Mux(ir_node *n) {
3242 ir_node *oldn = n, *sel = get_Mux_sel(n);
3243 ir_mode *mode = get_irn_mode(n);
3245 if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(mode)) {
3246 ir_node *cmp = get_Proj_pred(sel);
3247 long proj_nr = get_Proj_proj(sel);
3248 ir_node *f = get_Mux_false(n);
3249 ir_node *t = get_Mux_true(n);
3251 if (get_irn_op(cmp) == op_Cmp && classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
3252 ir_node *block = get_irn_n(n, -1);
3255 * Note: normalization puts the constant on the right site,
3256 * so we check only one case.
3258 * Note further that these optimization work even for floating point
3259 * with NaN's because -NaN == NaN.
3260 * However, if +0 and -0 is handled differently, we cannot use the first one.
3262 if (get_irn_op(f) == op_Minus &&
3263 get_Minus_op(f) == t &&
3264 get_Cmp_left(cmp) == t) {
3266 if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3267 /* Mux(a >=/> 0, -a, a) ==> Abs(a) */
3268 n = new_rd_Abs(get_irn_dbg_info(n),
3272 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3274 } else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3275 /* Mux(a <=/< 0, -a, a) ==> Minus(Abs(a)) */
3276 n = new_rd_Abs(get_irn_dbg_info(n),
3280 n = new_rd_Minus(get_irn_dbg_info(n),
3285 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3288 } else if (get_irn_op(t) == op_Minus &&
3289 get_Minus_op(t) == f &&
3290 get_Cmp_left(cmp) == f) {
3292 if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3293 /* Mux(a <=/< 0, a, -a) ==> Abs(a) */
3294 n = new_rd_Abs(get_irn_dbg_info(n),
3298 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3300 } else if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3301 /* Mux(a >=/> 0, a, -a) ==> Minus(Abs(a)) */
3302 n = new_rd_Abs(get_irn_dbg_info(n),
3306 n = new_rd_Minus(get_irn_dbg_info(n),
3311 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3316 if (mode_is_int(mode) && mode_is_signed(mode) &&
3317 get_mode_arithmetic(mode) == irma_twos_complement) {
3318 ir_node *x = get_Cmp_left(cmp);
3320 /* the following optimization works only with signed integer two-complement mode */
3322 if (mode == get_irn_mode(x)) {
3324 * FIXME: this restriction is two rigid, as it would still
3325 * work if mode(x) = Hs and mode == Is, but at least it removes
3328 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Le) &&
3329 classify_Const(t) == CNST_ALL_ONE &&
3330 classify_Const(f) == CNST_NULL) {
3332 * Mux(x:T </<= 0, 0, -1) -> Shrs(x, sizeof_bits(T) - 1)
3336 n = new_rd_Shrs(get_irn_dbg_info(n),
3337 current_ir_graph, block, x,
3338 new_r_Const_long(current_ir_graph, block, mode_Iu,
3339 get_mode_size_bits(mode) - 1),
3341 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3343 } else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) &&
3344 classify_Const(t) == CNST_ONE &&
3345 classify_Const(f) == CNST_NULL) {
3347 * Mux(x:T >/>= 0, 0, 1) -> Shr(-x, sizeof_bits(T) - 1)
3351 n = new_rd_Shr(get_irn_dbg_info(n),
3352 current_ir_graph, block,
3353 new_r_Minus(current_ir_graph, block, x, mode),
3354 new_r_Const_long(current_ir_graph, block, mode_Iu,
3355 get_mode_size_bits(mode) - 1),
3357 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3364 return arch_transform_node_Mux(n);
3365 } /* transform_node_Mux */
3368 * Optimize a Psi into some simpler cases.
3370 static ir_node *transform_node_Psi(ir_node *n) {
3372 return transform_node_Mux(n);
3375 } /* transform_node_Psi */
3378 * Tries several [inplace] [optimizing] transformations and returns an
3379 * equivalent node. The difference to equivalent_node() is that these
3380 * transformations _do_ generate new nodes, and thus the old node must
3381 * not be freed even if the equivalent node isn't the old one.
3383 static ir_node *transform_node(ir_node *n) {
3384 if (n->op->ops.transform_node)
3385 n = n->op->ops.transform_node(n);
3387 } /* transform_node */
3390 * Sets the default transform node operation for an ir_op_ops.
3392 * @param code the opcode for the default operation
3393 * @param ops the operations initialized
3398 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
3402 ops->transform_node = transform_node_##a; \
3435 } /* firm_set_default_transform_node */
3438 /* **************** Common Subexpression Elimination **************** */
3440 /** The size of the hash table used, should estimate the number of nodes
3442 #define N_IR_NODES 512
3444 /** Compares the attributes of two Const nodes. */
3445 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
3446 return (get_Const_tarval(a) != get_Const_tarval(b))
3447 || (get_Const_type(a) != get_Const_type(b));
3448 } /* node_cmp_attr_Const */
3450 /** Compares the attributes of two Proj nodes. */
3451 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
3452 return get_irn_proj_attr (a) != get_irn_proj_attr (b);
3453 } /* node_cmp_attr_Proj */
3455 /** Compares the attributes of two Filter nodes. */
3456 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
3457 return get_Filter_proj(a) != get_Filter_proj(b);
3458 } /* node_cmp_attr_Filter */
3460 /** Compares the attributes of two Alloc nodes. */
3461 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
3462 return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where)
3463 || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type);
3464 } /* node_cmp_attr_Alloc */
3466 /** Compares the attributes of two Free nodes. */
3467 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
3468 return (get_irn_free_attr(a).where != get_irn_free_attr(b).where)
3469 || (get_irn_free_attr(a).type != get_irn_free_attr(b).type);
3470 } /* node_cmp_attr_Free */
3472 /** Compares the attributes of two SymConst nodes. */
3473 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
3474 return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num)
3475 || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p)
3476 || (get_irn_symconst_attr(a).tp != get_irn_symconst_attr(b).tp);
3477 } /* node_cmp_attr_SymConst */
3479 /** Compares the attributes of two Call nodes. */
3480 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
3481 return (get_irn_call_attr(a) != get_irn_call_attr(b));
3482 } /* node_cmp_attr_Call */
3484 /** Compares the attributes of two Sel nodes. */
3485 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
3486 return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind)
3487 || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name)
3488 || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner)
3489 || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name)
3490 || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type);
3491 } /* node_cmp_attr_Sel */
3493 /** Compares the attributes of two Phi nodes. */
3494 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
3495 /* we can only enter this function if both nodes have the same number of inputs,
3496 hence it is enough to check if one of them is a Phi0 */
3498 /* check the Phi0 attribute */
3499 return get_irn_phi0_attr(a) != get_irn_phi0_attr(b);
3502 } /* node_cmp_attr_Phi */
3504 /** Compares the attributes of two Conv nodes. */
3505 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
3506 return get_Conv_strict(a) != get_Conv_strict(b);
3507 } /* node_cmp_attr_Conv */
3509 /** Compares the attributes of two Cast nodes. */
3510 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
3511 return get_Cast_type(a) != get_Cast_type(b);
3512 } /* node_cmp_attr_Cast */
3514 /** Compares the attributes of two Load nodes. */
3515 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
3516 if (get_Load_volatility(a) == volatility_is_volatile ||
3517 get_Load_volatility(b) == volatility_is_volatile)
3518 /* NEVER do CSE on volatile Loads */
3521 return get_Load_mode(a) != get_Load_mode(b);
3522 } /* node_cmp_attr_Load */
3524 /** Compares the attributes of two Store nodes. */
3525 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
3526 /* NEVER do CSE on volatile Stores */
3527 return (get_Store_volatility(a) == volatility_is_volatile ||
3528 get_Store_volatility(b) == volatility_is_volatile);
3529 } /* node_cmp_attr_Store */
3531 /** Compares the attributes of two Confirm nodes. */
3532 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
3533 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
3534 } /* node_cmp_attr_Confirm */
3536 /** Compares the attributes of two ASM nodes. */
3537 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
3539 const ir_asm_constraint *ca;
3540 const ir_asm_constraint *cb;
3543 if (get_ASM_text(a) != get_ASM_text(b));
3546 /* Should we really check the constraints here? Should be better, but is strange. */
3547 n = get_ASM_n_input_constraints(a);
3548 if (n != get_ASM_n_input_constraints(b))
3551 ca = get_ASM_input_constraints(a);
3552 cb = get_ASM_input_constraints(b);
3553 for (i = 0; i < n; ++i) {
3554 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
3558 n = get_ASM_n_output_constraints(a);
3559 if (n != get_ASM_n_output_constraints(b))
3562 ca = get_ASM_output_constraints(a);
3563 cb = get_ASM_output_constraints(b);
3564 for (i = 0; i < n; ++i) {
3565 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
3569 n = get_ASM_n_clobbers(a);
3570 if (n != get_ASM_n_clobbers(b))
3573 cla = get_ASM_clobbers(a);
3574 clb = get_ASM_clobbers(b);
3575 for (i = 0; i < n; ++i) {
3576 if (cla[i] != clb[i])
3580 } /* node_cmp_attr_ASM */
3583 * Set the default node attribute compare operation for an ir_op_ops.
3585 * @param code the opcode for the default operation
3586 * @param ops the operations initialized
3591 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
3595 ops->node_cmp_attr = node_cmp_attr_##a; \
3620 } /* firm_set_default_node_cmp_attr */
3623 * Compare function for two nodes in the hash table. Gets two
3624 * nodes as parameters. Returns 0 if the nodes are a cse.
3626 int identities_cmp(const void *elt, const void *key) {
3633 if (a == b) return 0;
3635 if ((get_irn_op(a) != get_irn_op(b)) ||
3636 (get_irn_mode(a) != get_irn_mode(b))) return 1;
3638 /* compare if a's in and b's in are of equal length */
3639 irn_arity_a = get_irn_intra_arity (a);
3640 if (irn_arity_a != get_irn_intra_arity(b))
3643 /* for block-local cse and op_pin_state_pinned nodes: */
3644 if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) {
3645 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
3649 /* compare a->in[0..ins] with b->in[0..ins] */
3650 for (i = 0; i < irn_arity_a; i++)
3651 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
3655 * here, we already now that the nodes are identical except their
3658 if (a->op->ops.node_cmp_attr)
3659 return a->op->ops.node_cmp_attr(a, b);
3662 } /* identities_cmp */
3665 * Calculate a hash value of a node.
3667 unsigned ir_node_hash(ir_node *node) {
3671 if (node->op == op_Const) {
3672 /* special value for const, as they only differ in their tarval. */
3673 h = HASH_PTR(node->attr.con.tv);
3674 h = 9*h + HASH_PTR(get_irn_mode(node));
3675 } else if (node->op == op_SymConst) {
3676 /* special value for const, as they only differ in their symbol. */
3677 h = HASH_PTR(node->attr.symc.sym.type_p);
3678 h = 9*h + HASH_PTR(get_irn_mode(node));
3681 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
3682 h = irn_arity = get_irn_intra_arity(node);
3684 /* consider all in nodes... except the block if not a control flow. */
3685 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
3686 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
3690 h = 9*h + HASH_PTR(get_irn_mode(node));
3692 h = 9*h + HASH_PTR(get_irn_op(node));
3696 } /* ir_node_hash */
3698 pset *new_identities(void) {
3699 return new_pset(identities_cmp, N_IR_NODES);
3700 } /* new_identities */
3702 void del_identities(pset *value_table) {
3703 del_pset(value_table);
3704 } /* del_identities */
3707 * Return the canonical node computing the same value as n.
3709 * @param value_table The value table
3710 * @param n The node to lookup
3712 * Looks up the node in a hash table.
3714 * For Const nodes this is performed in the constructor, too. Const
3715 * nodes are extremely time critical because of their frequent use in
3716 * constant string arrays.
3718 static INLINE ir_node *identify(pset *value_table, ir_node *n) {
3721 if (!value_table) return n;
3723 if (get_opt_reassociation()) {
3724 if (is_op_commutative(get_irn_op(n))) {
3725 ir_node *l = get_binop_left(n);
3726 ir_node *r = get_binop_right(n);
3728 /* for commutative operators perform a OP b == b OP a */
3729 if (get_irn_idx(l) > get_irn_idx(r)) {
3730 set_binop_left(n, r);
3731 set_binop_right(n, l);
3736 o = pset_find(value_table, n, ir_node_hash(n));
3745 * During construction we set the op_pin_state_pinned flag in the graph right when the
3746 * optimization is performed. The flag turning on procedure global cse could
3747 * be changed between two allocations. This way we are safe.
3749 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
3752 n = identify(value_table, n);
3753 if (get_irn_n(old, -1) != get_irn_n(n, -1))
3754 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3756 } /* identify_cons */
3759 * Return the canonical node computing the same value as n.
3760 * Looks up the node in a hash table, enters it in the table
3761 * if it isn't there yet.
3763 ir_node *identify_remember(pset *value_table, ir_node *n) {
3766 if (!value_table) return n;
3768 if (get_opt_reassociation()) {
3769 if (is_op_commutative(get_irn_op(n))) {
3770 ir_node *l = get_binop_left(n);
3771 ir_node *r = get_binop_right(n);
3772 int l_idx = get_irn_idx(l);
3773 int r_idx = get_irn_idx(r);
3775 /* For commutative operators perform a OP b == b OP a but keep
3776 constants on the RIGHT side. This helps greatly in some optimizations.
3777 Moreover we use the idx number to make the form deterministic. */
3778 if (is_irn_constlike(l))
3780 if (is_irn_constlike(r))
3782 if (l_idx < r_idx) {
3783 set_binop_left(n, r);
3784 set_binop_right(n, l);
3789 /* lookup or insert in hash table with given hash key. */
3790 o = pset_insert(value_table, n, ir_node_hash(n));
3797 } /* identify_remember */
3799 /* Add a node to the identities value table. */
3800 void add_identities(pset *value_table, ir_node *node) {
3801 if (get_opt_cse() && is_no_Block(node))
3802 identify_remember(value_table, node);
3803 } /* add_identities */
3805 /* Visit each node in the value table of a graph. */
3806 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
3808 ir_graph *rem = current_ir_graph;
3810 current_ir_graph = irg;
3811 foreach_pset(irg->value_table, node)
3813 current_ir_graph = rem;
3814 } /* visit_all_identities */
3817 * Garbage in, garbage out. If a node has a dead input, i.e., the
3818 * Bad node is input to the node, return the Bad node.
3820 static INLINE ir_node *gigo(ir_node *node) {
3822 ir_op *op = get_irn_op(node);
3824 /* remove garbage blocks by looking at control flow that leaves the block
3825 and replacing the control flow by Bad. */
3826 if (get_irn_mode(node) == mode_X) {
3827 ir_node *block = get_nodes_block(skip_Proj(node));
3829 /* Don't optimize nodes in immature blocks. */
3830 if (!get_Block_matured(block)) return node;
3831 /* Don't optimize End, may have Bads. */
3832 if (op == op_End) return node;
3834 if (is_Block(block)) {
3835 irn_arity = get_irn_arity(block);
3836 for (i = 0; i < irn_arity; i++) {
3837 if (!is_Bad(get_irn_n(block, i)))
3840 if (i == irn_arity) {
3841 ir_graph *irg = get_irn_irg(block);
3842 /* the start block is never dead */
3843 if (block != get_irg_start_block(irg)
3844 && block != get_irg_end_block(irg))
3850 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
3851 blocks predecessors is dead. */
3852 if (op != op_Block && op != op_Phi && op != op_Tuple) {
3853 irn_arity = get_irn_arity(node);
3856 * Beware: we can only read the block of a non-floating node.
3858 if (is_irn_pinned_in_irg(node) &&
3859 is_Block_dead(get_nodes_block(node)))
3862 for (i = 0; i < irn_arity; i++) {
3863 ir_node *pred = get_irn_n(node, i);
3868 /* Propagating Unknowns here seems to be a bad idea, because
3869 sometimes we need a node as a input and did not want that
3871 However, it might be useful to move this into a later phase
3872 (if you think that optimizing such code is useful). */
3873 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
3874 return new_Unknown(get_irn_mode(node));
3879 /* With this code we violate the agreement that local_optimize
3880 only leaves Bads in Block, Phi and Tuple nodes. */
3881 /* If Block has only Bads as predecessors it's garbage. */
3882 /* If Phi has only Bads as predecessors it's garbage. */
3883 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
3884 irn_arity = get_irn_arity(node);
3885 for (i = 0; i < irn_arity; i++) {
3886 if (!is_Bad(get_irn_n(node, i))) break;
3888 if (i == irn_arity) node = new_Bad();
3895 * These optimizations deallocate nodes from the obstack.
3896 * It can only be called if it is guaranteed that no other nodes
3897 * reference this one, i.e., right after construction of a node.
3899 * @param n The node to optimize
3901 * current_ir_graph must be set to the graph of the node!
3903 ir_node *optimize_node(ir_node *n) {
3906 ir_opcode iro = get_irn_opcode(n);
3908 /* Always optimize Phi nodes: part of the construction. */
3909 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
3911 /* constant expression evaluation / constant folding */
3912 if (get_opt_constant_folding()) {
3913 /* neither constants nor Tuple values can be evaluated */
3914 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
3915 /* try to evaluate */
3916 tv = computed_value(n);
3917 if (tv != tarval_bad) {
3919 ir_type *old_tp = get_irn_type(n);
3920 int i, arity = get_irn_arity(n);
3924 * Try to recover the type of the new expression.
3926 for (i = 0; i < arity && !old_tp; ++i)
3927 old_tp = get_irn_type(get_irn_n(n, i));
3930 * we MUST copy the node here temporary, because it's still needed
3931 * for DBG_OPT_CSTEVAL
3933 node_size = offsetof(ir_node, attr) + n->op->attr_size;
3934 oldn = alloca(node_size);
3936 memcpy(oldn, n, node_size);
3937 CLONE_ARR_A(ir_node *, oldn->in, n->in);
3939 /* ARG, copy the in array, we need it for statistics */
3940 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
3942 /* note the inplace edges module */
3943 edges_node_deleted(n, current_ir_graph);
3945 /* evaluation was successful -- replace the node. */
3946 irg_kill_node(current_ir_graph, n);
3947 nw = new_Const(get_tarval_mode (tv), tv);
3949 if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv))
3950 set_Const_type(nw, old_tp);
3951 DBG_OPT_CSTEVAL(oldn, nw);
3957 /* remove unnecessary nodes */
3958 if (get_opt_constant_folding() ||
3959 (iro == iro_Phi) || /* always optimize these nodes. */
3961 (iro == iro_Proj) ||
3962 (iro == iro_Block) ) /* Flags tested local. */
3963 n = equivalent_node(n);
3965 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
3967 /* Common Subexpression Elimination.
3969 * Checks whether n is already available.
3970 * The block input is used to distinguish different subexpressions. Right
3971 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
3972 * subexpressions within a block.
3975 n = identify_cons(current_ir_graph->value_table, n);
3978 edges_node_deleted(oldn, current_ir_graph);
3980 /* We found an existing, better node, so we can deallocate the old node. */
3981 irg_kill_node(current_ir_graph, oldn);
3985 /* Some more constant expression evaluation that does not allow to
3987 iro = get_irn_opcode(n);
3988 if (get_opt_constant_folding() ||
3989 (iro == iro_Cond) ||
3990 (iro == iro_Proj)) /* Flags tested local. */
3991 n = transform_node(n);
3993 /* Remove nodes with dead (Bad) input.
3994 Run always for transformation induced Bads. */
3997 /* Now we have a legal, useful node. Enter it in hash table for CSE */
3998 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
3999 n = identify_remember(current_ir_graph->value_table, n);
4003 } /* optimize_node */
4007 * These optimizations never deallocate nodes (in place). This can cause dead
4008 * nodes lying on the obstack. Remove these by a dead node elimination,
4009 * i.e., a copying garbage collection.
4011 ir_node *optimize_in_place_2(ir_node *n) {
4014 ir_opcode iro = get_irn_opcode(n);
4016 if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n;
4018 /* constant expression evaluation / constant folding */
4019 if (get_opt_constant_folding()) {
4020 /* neither constants nor Tuple values can be evaluated */
4021 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
4022 /* try to evaluate */
4023 tv = computed_value(n);
4024 if (tv != tarval_bad) {
4025 /* evaluation was successful -- replace the node. */
4026 ir_type *old_tp = get_irn_type(n);
4027 int i, arity = get_irn_arity(n);
4030 * Try to recover the type of the new expression.
4032 for (i = 0; i < arity && !old_tp; ++i)
4033 old_tp = get_irn_type(get_irn_n(n, i));
4035 n = new_Const(get_tarval_mode(tv), tv);
4037 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
4038 set_Const_type(n, old_tp);
4040 DBG_OPT_CSTEVAL(oldn, n);
4046 /* remove unnecessary nodes */
4047 if (get_opt_constant_folding() ||
4048 (iro == iro_Phi) || /* always optimize these nodes. */
4049 (iro == iro_Id) || /* ... */
4050 (iro == iro_Proj) || /* ... */
4051 (iro == iro_Block) ) /* Flags tested local. */
4052 n = equivalent_node(n);
4054 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
4056 /** common subexpression elimination **/
4057 /* Checks whether n is already available. */
4058 /* The block input is used to distinguish different subexpressions. Right
4059 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
4060 subexpressions within a block. */
4061 if (get_opt_cse()) {
4062 n = identify(current_ir_graph->value_table, n);
4065 /* Some more constant expression evaluation. */
4066 iro = get_irn_opcode(n);
4067 if (get_opt_constant_folding() ||
4068 (iro == iro_Cond) ||
4069 (iro == iro_Proj)) /* Flags tested local. */
4070 n = transform_node(n);
4072 /* Remove nodes with dead (Bad) input.
4073 Run always for transformation induced Bads. */
4076 /* Now we can verify the node, as it has no dead inputs any more. */
4079 /* Now we have a legal, useful node. Enter it in hash table for cse.
4080 Blocks should be unique anyways. (Except the successor of start:
4081 is cse with the start block!) */
4082 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
4083 n = identify_remember(current_ir_graph->value_table, n);
4086 } /* optimize_in_place_2 */
4089 * Wrapper for external use, set proper status bits after optimization.
4091 ir_node *optimize_in_place(ir_node *n) {
4092 /* Handle graph state */
4093 assert(get_irg_phase_state(current_ir_graph) != phase_building);
4095 if (get_opt_global_cse())
4096 set_irg_pinned(current_ir_graph, op_pin_state_floats);
4097 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
4098 set_irg_outs_inconsistent(current_ir_graph);
4100 /* FIXME: Maybe we could also test whether optimizing the node can
4101 change the control graph. */
4102 set_irg_doms_inconsistent(current_ir_graph);
4103 return optimize_in_place_2(n);
4104 } /* optimize_in_place */
4107 * Sets the default operation for an ir_ops.
4109 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
4110 ops = firm_set_default_computed_value(code, ops);
4111 ops = firm_set_default_equivalent_node(code, ops);
4112 ops = firm_set_default_transform_node(code, ops);
4113 ops = firm_set_default_node_cmp_attr(code, ops);
4114 ops = firm_set_default_get_type(code, ops);
4115 ops = firm_set_default_get_type_attr(code, ops);
4116 ops = firm_set_default_get_entity_attr(code, ops);
4119 } /* firm_set_default_operations */