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) {
670 ops->computed_value = computed_value_##a; \
705 } /* firm_set_default_computed_value */
708 * Returns a equivalent block for another block.
709 * If the block has only one predecessor, this is
710 * the equivalent one. If the only predecessor of a block is
711 * the block itself, this is a dead block.
713 * If both predecessors of a block are the branches of a binary
714 * Cond, the equivalent block is Cond's block.
716 * If all predecessors of a block are bad or lies in a dead
717 * block, the current block is dead as well.
719 * Note, that blocks are NEVER turned into Bad's, instead
720 * the dead_block flag is set. So, never test for is_Bad(block),
721 * always use is_dead_Block(block).
723 static ir_node *equivalent_node_Block(ir_node *n)
726 int n_preds = get_Block_n_cfgpreds(n);
728 /* The Block constructor does not call optimize, but mature_immBlock
729 calls the optimization. */
730 assert(get_Block_matured(n));
732 /* Straightening: a single entry Block following a single exit Block
733 can be merged, if it is not the Start block. */
734 /* !!! Beware, all Phi-nodes of n must have been optimized away.
735 This should be true, as the block is matured before optimize is called.
736 But what about Phi-cycles with the Phi0/Id that could not be resolved?
737 Remaining Phi nodes are just Ids. */
738 if ((n_preds == 1) && (get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp)) {
739 ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0));
740 if (predblock == oldn) {
741 /* Jmp jumps into the block it is in -- deal self cycle. */
742 n = set_Block_dead(n);
743 DBG_OPT_DEAD_BLOCK(oldn, n);
744 } else if (get_opt_control_flow_straightening()) {
746 DBG_OPT_STG(oldn, n);
748 } else if ((n_preds == 1) &&
749 (get_irn_op(skip_Proj(get_Block_cfgpred(n, 0))) == op_Cond)) {
750 ir_node *predblock = get_Block_cfgpred_block(n, 0);
751 if (predblock == oldn) {
752 /* Jmp jumps into the block it is in -- deal self cycle. */
753 n = set_Block_dead(n);
754 DBG_OPT_DEAD_BLOCK(oldn, n);
756 } else if ((n_preds == 2) &&
757 (get_opt_control_flow_weak_simplification())) {
758 /* Test whether Cond jumps twice to this block
759 * The more general case which more than 2 predecessors is handles
760 * in optimize_cf(), we handle only this special case for speed here.
762 ir_node *a = get_Block_cfgpred(n, 0);
763 ir_node *b = get_Block_cfgpred(n, 1);
765 if ((get_irn_op(a) == op_Proj) &&
766 (get_irn_op(b) == op_Proj) &&
767 (get_Proj_pred(a) == get_Proj_pred(b)) &&
768 (get_irn_op(get_Proj_pred(a)) == op_Cond) &&
769 (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) {
770 /* Also a single entry Block following a single exit Block. Phis have
771 twice the same operand and will be optimized away. */
772 n = get_nodes_block(get_Proj_pred(a));
773 DBG_OPT_IFSIM1(oldn, a, b, n);
775 } else if (get_opt_unreachable_code() &&
776 (n != get_irg_start_block(current_ir_graph)) &&
777 (n != get_irg_end_block(current_ir_graph)) ) {
780 /* If all inputs are dead, this block is dead too, except if it is
781 the start or end block. This is one step of unreachable code
783 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
784 ir_node *pred = get_Block_cfgpred(n, i);
787 if (is_Bad(pred)) continue;
788 pred_blk = get_nodes_block(skip_Proj(pred));
790 if (is_Block_dead(pred_blk)) continue;
793 /* really found a living input */
798 n = set_Block_dead(n);
799 DBG_OPT_DEAD_BLOCK(oldn, n);
804 } /* equivalent_node_Block */
807 * Returns a equivalent node for a Jmp, a Bad :-)
808 * Of course this only happens if the Block of the Jmp is dead.
810 static ir_node *equivalent_node_Jmp(ir_node *n) {
811 /* unreachable code elimination */
812 if (is_Block_dead(get_nodes_block(n)))
816 } /* equivalent_node_Jmp */
818 /** Raise is handled in the same way as Jmp. */
819 #define equivalent_node_Raise equivalent_node_Jmp
822 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
823 See transform_node_Proj_Cond(). */
826 * Optimize operations that are commutative and have neutral 0,
827 * so a op 0 = 0 op a = a.
829 static ir_node *equivalent_node_neutral_zero(ir_node *n)
833 ir_node *a = get_binop_left(n);
834 ir_node *b = get_binop_right(n);
839 /* After running compute_node there is only one constant predecessor.
840 Find this predecessors value and remember the other node: */
841 if ((tv = value_of(a)) != tarval_bad) {
843 } else if ((tv = value_of(b)) != tarval_bad) {
848 /* If this predecessors constant value is zero, the operation is
849 * unnecessary. Remove it.
851 * Beware: If n is a Add, the mode of on and n might be different
852 * which happens in this rare construction: NULL + 3.
853 * Then, a Conv would be needed which we cannot include here.
855 if (classify_tarval (tv) == TV_CLASSIFY_NULL) {
856 if (get_irn_mode(on) == get_irn_mode(n)) {
859 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
864 } /* equivalent_node_neutral_zero */
867 * Eor is commutative and has neutral 0.
869 #define equivalent_node_Eor equivalent_node_neutral_zero
872 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
874 * The second one looks strange, but this construct
875 * is used heavily in the LCC sources :-).
877 * Beware: The Mode of an Add may be different than the mode of its
878 * predecessors, so we could not return a predecessors in all cases.
880 static ir_node *equivalent_node_Add(ir_node *n) {
882 ir_node *left, *right;
883 ir_mode *mode = get_irn_mode(n);
885 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
886 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
889 n = equivalent_node_neutral_zero(n);
893 left = get_Add_left(n);
894 right = get_Add_right(n);
896 if (get_irn_op(left) == op_Sub) {
897 if (get_Sub_right(left) == right) {
900 n = get_Sub_left(left);
901 if (mode == get_irn_mode(n)) {
902 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
907 if (get_irn_op(right) == op_Sub) {
908 if (get_Sub_right(right) == left) {
911 n = get_Sub_left(right);
912 if (mode == get_irn_mode(n)) {
913 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
919 } /* equivalent_node_Add */
922 * optimize operations that are not commutative but have neutral 0 on left,
925 static ir_node *equivalent_node_left_zero(ir_node *n) {
928 ir_node *a = get_binop_left(n);
929 ir_node *b = get_binop_right(n);
931 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
934 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
937 } /* equivalent_node_left_zero */
939 #define equivalent_node_Shl equivalent_node_left_zero
940 #define equivalent_node_Shr equivalent_node_left_zero
941 #define equivalent_node_Shrs equivalent_node_left_zero
942 #define equivalent_node_Rot equivalent_node_left_zero
945 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
947 * The second one looks strange, but this construct
948 * is used heavily in the LCC sources :-).
950 * Beware: The Mode of a Sub may be different than the mode of its
951 * predecessors, so we could not return a predecessors in all cases.
953 static ir_node *equivalent_node_Sub(ir_node *n) {
956 ir_mode *mode = get_irn_mode(n);
958 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
959 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
962 b = get_Sub_right(n);
964 /* Beware: modes might be different */
965 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
966 ir_node *a = get_Sub_left(n);
967 if (mode == get_irn_mode(a)) {
970 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
974 } /* equivalent_node_Sub */
978 * Optimize an "idempotent unary op", ie op(op(n)) = n.
981 * -(-a) == a, but might overflow two times.
982 * We handle it anyway here but the better way would be a
983 * flag. This would be needed for Pascal for instance.
985 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
987 ir_node *pred = get_unop_op(n);
989 /* optimize symmetric unop */
990 if (get_irn_op(pred) == get_irn_op(n)) {
991 n = get_unop_op(pred);
992 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
995 } /* equivalent_node_idempotent_unop */
997 /** Optimize Not(Not(x)) == x. */
998 #define equivalent_node_Not equivalent_node_idempotent_unop
1000 /** -(-x) == x ??? Is this possible or can --x raise an
1001 out of bounds exception if min =! max? */
1002 #define equivalent_node_Minus equivalent_node_idempotent_unop
1005 * Optimize a * 1 = 1 * a = a.
1007 static ir_node *equivalent_node_Mul(ir_node *n) {
1009 ir_node *a = get_Mul_left(n);
1010 ir_node *b = get_Mul_right(n);
1012 /* Mul is commutative and has again an other neutral element. */
1013 if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) {
1015 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1016 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) {
1018 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1021 } /* equivalent_node_Mul */
1024 * Optimize a / 1 = a.
1026 static ir_node *equivalent_node_Div(ir_node *n) {
1027 ir_node *a = get_Div_left(n);
1028 ir_node *b = get_Div_right(n);
1030 /* Div is not commutative. */
1031 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1032 /* Turn Div into a tuple (mem, bad, a) */
1033 ir_node *mem = get_Div_mem(n);
1034 ir_node *blk = get_irn_n(n, -1);
1035 turn_into_tuple(n, pn_Div_max);
1036 set_Tuple_pred(n, pn_Div_M, mem);
1037 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
1038 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
1039 set_Tuple_pred(n, pn_Div_res, a);
1042 } /* equivalent_node_Div */
1045 * Optimize a / 1.0 = a.
1047 static ir_node *equivalent_node_Quot(ir_node *n) {
1048 ir_node *a = get_Quot_left(n);
1049 ir_node *b = get_Quot_right(n);
1051 /* Div is not commutative. */
1052 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* Quot(x, 1) == x */
1053 /* Turn Quot into a tuple (mem, jmp, bad, a) */
1054 ir_node *mem = get_Quot_mem(n);
1055 ir_node *blk = get_irn_n(n, -1);
1056 turn_into_tuple(n, pn_Quot_max);
1057 set_Tuple_pred(n, pn_Quot_M, mem);
1058 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
1059 set_Tuple_pred(n, pn_Quot_X_except, new_Bad()); /* no exception */
1060 set_Tuple_pred(n, pn_Quot_res, a);
1063 } /* equivalent_node_Quot */
1066 * Optimize a / 1 = a.
1068 static ir_node *equivalent_node_DivMod(ir_node *n) {
1069 ir_node *b = get_DivMod_right(n);
1071 /* Div is not commutative. */
1072 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1073 /* Turn DivMod into a tuple (mem, jmp, bad, a, 0) */
1074 ir_node *a = get_DivMod_left(n);
1075 ir_node *mem = get_Div_mem(n);
1076 ir_node *blk = get_irn_n(n, -1);
1077 ir_mode *mode = get_DivMod_resmode(n);
1079 turn_into_tuple(n, pn_DivMod_max);
1080 set_Tuple_pred(n, pn_DivMod_M, mem);
1081 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
1082 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1083 set_Tuple_pred(n, pn_DivMod_res_div, a);
1084 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1087 } /* equivalent_node_DivMod */
1090 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1092 static ir_node *equivalent_node_Or(ir_node *n) {
1095 ir_node *a = get_Or_left(n);
1096 ir_node *b = get_Or_right(n);
1099 n = a; /* Or has it's own neutral element */
1100 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1101 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) {
1103 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1104 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
1106 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1110 } /* equivalent_node_Or */
1113 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a.
1115 static ir_node *equivalent_node_And(ir_node *n) {
1118 ir_node *a = get_And_left(n);
1119 ir_node *b = get_And_right(n);
1122 n = a; /* And has it's own neutral element */
1123 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1124 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) {
1126 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1127 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) {
1129 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1132 } /* equivalent_node_And */
1135 * Try to remove useless Conv's:
1137 static ir_node *equivalent_node_Conv(ir_node *n) {
1139 ir_node *a = get_Conv_op(n);
1142 ir_mode *n_mode = get_irn_mode(n);
1143 ir_mode *a_mode = get_irn_mode(a);
1145 if (n_mode == a_mode) { /* No Conv necessary */
1146 /* leave strict floating point Conv's */
1147 if (get_Conv_strict(n))
1150 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1151 } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */
1155 n_mode = get_irn_mode(n);
1156 b_mode = get_irn_mode(b);
1158 if (n_mode == b_mode) {
1159 if (n_mode == mode_b) {
1160 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1161 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1162 } else if (mode_is_int(n_mode)) {
1163 if (smaller_mode(b_mode, a_mode)){
1164 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1165 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1171 } /* equivalent_node_Conv */
1174 * A Cast may be removed if the type of the previous node
1175 * is already the type of the Cast.
1177 static ir_node *equivalent_node_Cast(ir_node *n) {
1179 ir_node *pred = get_Cast_op(n);
1181 if (get_irn_type(pred) == get_Cast_type(n)) {
1183 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1186 } /* equivalent_node_Cast */
1189 * Several optimizations:
1190 * - no Phi in start block.
1191 * - remove Id operators that are inputs to Phi
1192 * - fold Phi-nodes, iff they have only one predecessor except
1195 static ir_node *equivalent_node_Phi(ir_node *n) {
1199 ir_node *block = NULL; /* to shutup gcc */
1200 ir_node *first_val = NULL; /* to shutup gcc */
1202 if (!get_opt_normalize()) return n;
1204 n_preds = get_Phi_n_preds(n);
1206 block = get_nodes_block(n);
1207 /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
1208 assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
1209 if ((is_Block_dead(block)) || /* Control dead */
1210 (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */
1211 return new_Bad(); /* in the Start Block. */
1213 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1215 /* If the Block has a Bad pred, we also have one. */
1216 for (i = 0; i < n_preds; ++i)
1217 if (is_Bad(get_Block_cfgpred(block, i)))
1218 set_Phi_pred(n, i, new_Bad());
1220 /* Find first non-self-referencing input */
1221 for (i = 0; i < n_preds; ++i) {
1222 first_val = get_Phi_pred(n, i);
1223 if ( (first_val != n) /* not self pointer */
1225 && (! is_Bad(first_val))
1227 ) { /* value not dead */
1228 break; /* then found first value. */
1233 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1237 /* search for rest of inputs, determine if any of these
1238 are non-self-referencing */
1239 while (++i < n_preds) {
1240 ir_node *scnd_val = get_Phi_pred(n, i);
1241 if ( (scnd_val != n)
1242 && (scnd_val != first_val)
1244 && (! is_Bad(scnd_val))
1252 /* Fold, if no multiple distinct non-self-referencing inputs */
1254 DBG_OPT_PHI(oldn, n);
1257 } /* equivalent_node_Phi */
1260 * Several optimizations:
1261 * - no Sync in start block.
1262 * - fold Sync-nodes, iff they have only one predecessor except
1265 static ir_node *equivalent_node_Sync(ir_node *n) {
1269 ir_node *first_val = NULL; /* to shutup gcc */
1271 if (!get_opt_normalize()) return n;
1273 n_preds = get_Sync_n_preds(n);
1275 /* Find first non-self-referencing input */
1276 for (i = 0; i < n_preds; ++i) {
1277 first_val = get_Sync_pred(n, i);
1278 if ((first_val != n) /* not self pointer */ &&
1279 (! is_Bad(first_val))
1280 ) { /* value not dead */
1281 break; /* then found first value. */
1286 /* A totally Bad or self-referencing Sync (we didn't break the above loop) */
1289 /* search the rest of inputs, determine if any of these
1290 are non-self-referencing */
1291 while (++i < n_preds) {
1292 ir_node *scnd_val = get_Sync_pred(n, i);
1293 if ((scnd_val != n) &&
1294 (scnd_val != first_val) &&
1295 (! is_Bad(scnd_val))
1301 /* Fold, if no multiple distinct non-self-referencing inputs */
1303 DBG_OPT_SYNC(oldn, n);
1306 } /* equivalent_node_Sync */
1309 * Optimize Proj(Tuple) and gigo() for ProjX in Bad block,
1310 * ProjX(Load) and ProjX(Store).
1312 static ir_node *equivalent_node_Proj(ir_node *proj) {
1313 ir_node *oldn = proj;
1314 ir_node *a = get_Proj_pred(proj);
1316 if (get_irn_op(a) == op_Tuple) {
1317 /* Remove the Tuple/Proj combination. */
1318 if ( get_Proj_proj(proj) <= get_Tuple_n_preds(a) ) {
1319 proj = get_Tuple_pred(a, get_Proj_proj(proj));
1320 DBG_OPT_TUPLE(oldn, a, proj);
1322 /* This should not happen! */
1323 assert(! "found a Proj with higher number than Tuple predecessors");
1326 } else if (get_irn_mode(proj) == mode_X) {
1327 if (is_Block_dead(get_nodes_block(skip_Proj(proj)))) {
1328 /* Remove dead control flow -- early gigo(). */
1330 } else if (get_opt_ldst_only_null_ptr_exceptions()) {
1331 ir_op *op = get_irn_op(a);
1333 if (op == op_Load) {
1334 /* get the Load address */
1335 ir_node *addr = get_Load_ptr(a);
1336 ir_node *blk = get_irn_n(a, -1);
1339 if (value_not_null(addr, &confirm)) {
1340 if (confirm == NULL) {
1341 /* this node may float if it did not depend on a Confirm */
1342 set_irn_pinned(a, op_pin_state_floats);
1344 if (get_Proj_proj(proj) == pn_Load_X_except) {
1345 DBG_OPT_EXC_REM(proj);
1348 return new_r_Jmp(current_ir_graph, blk);
1350 } else if (op == op_Store) {
1351 /* get the load/store address */
1352 ir_node *addr = get_Store_ptr(a);
1353 ir_node *blk = get_irn_n(a, -1);
1356 if (value_not_null(addr, &confirm)) {
1357 if (confirm == NULL) {
1358 /* this node may float if it did not depend on a Confirm */
1359 set_irn_pinned(a, op_pin_state_floats);
1361 if (get_Proj_proj(proj) == pn_Store_X_except) {
1362 DBG_OPT_EXC_REM(proj);
1365 return new_r_Jmp(current_ir_graph, blk);
1372 } /* equivalent_node_Proj */
1377 static ir_node *equivalent_node_Id(ir_node *n) {
1382 } while (get_irn_op(n) == op_Id);
1384 DBG_OPT_ID(oldn, n);
1386 } /* equivalent_node_Id */
1391 static ir_node *equivalent_node_Mux(ir_node *n)
1393 ir_node *oldn = n, *sel = get_Mux_sel(n);
1394 tarval *ts = value_of(sel);
1396 /* Mux(true, f, t) == t */
1397 if (ts == tarval_b_true) {
1398 n = get_Mux_true(n);
1399 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1401 /* Mux(false, f, t) == f */
1402 else if (ts == tarval_b_false) {
1403 n = get_Mux_false(n);
1404 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1406 /* Mux(v, x, x) == x */
1407 else if (get_Mux_false(n) == get_Mux_true(n)) {
1408 n = get_Mux_true(n);
1409 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1411 else if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(get_irn_mode(n))) {
1412 ir_node *cmp = get_Proj_pred(sel);
1413 long proj_nr = get_Proj_proj(sel);
1414 ir_node *b = get_Mux_false(n);
1415 ir_node *a = get_Mux_true(n);
1418 * Note: normalization puts the constant on the right site,
1419 * so we check only one case.
1421 * Note further that these optimization work even for floating point
1422 * with NaN's because -NaN == NaN.
1423 * However, if +0 and -0 is handled differently, we cannot use the first one.
1425 if (get_irn_op(cmp) == op_Cmp && get_Cmp_left(cmp) == a) {
1426 if (classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
1427 /* Mux(a CMP 0, X, a) */
1428 if (get_irn_op(b) == op_Minus && get_Minus_op(b) == a) {
1429 /* Mux(a CMP 0, -a, a) */
1430 if (proj_nr == pn_Cmp_Eq) {
1431 /* Mux(a == 0, -a, a) ==> -a */
1433 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1434 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1435 /* Mux(a != 0, -a, a) ==> a */
1437 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1439 } else if (classify_Const(b) == CNST_NULL) {
1440 /* Mux(a CMP 0, 0, a) */
1441 if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1442 /* Mux(a != 0, 0, a) ==> a */
1444 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1445 } else if (proj_nr == pn_Cmp_Eq) {
1446 /* Mux(a == 0, 0, a) ==> 0 */
1448 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1455 } /* equivalent_node_Mux */
1458 * Returns a equivalent node of a Psi: if a condition is true
1459 * and all previous conditions are false we know its value.
1460 * If all conditions are false its value is the default one.
1462 static ir_node *equivalent_node_Psi(ir_node *n) {
1464 return equivalent_node_Mux(n);
1466 } /* equivalent_node_Psi */
1469 * Optimize -a CMP -b into b CMP a.
1470 * This works only for for modes where unary Minus
1472 * Note that two-complement integers can Overflow
1473 * so it will NOT work.
1475 * For == and != can be handled in Proj(Cmp)
1477 static ir_node *equivalent_node_Cmp(ir_node *n) {
1478 ir_node *left = get_Cmp_left(n);
1479 ir_node *right = get_Cmp_right(n);
1481 if (get_irn_op(left) == op_Minus && get_irn_op(right) == op_Minus &&
1482 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1483 left = get_Minus_op(left);
1484 right = get_Minus_op(right);
1485 set_Cmp_left(n, right);
1486 set_Cmp_right(n, left);
1489 } /* equivalent_node_Cmp */
1492 * Remove Confirm nodes if setting is on.
1493 * Replace Confirms(x, '=', Constlike) by Constlike.
1495 static ir_node *equivalent_node_Confirm(ir_node *n) {
1496 ir_node *pred = get_Confirm_value(n);
1497 pn_Cmp pnc = get_Confirm_cmp(n);
1499 if (get_irn_op(pred) == op_Confirm && pnc == get_Confirm_cmp(pred)) {
1501 * rare case: two identical Confirms one after another,
1502 * replace the second one with the first.
1506 if (pnc == pn_Cmp_Eq) {
1507 ir_node *bound = get_Confirm_bound(n);
1510 * Optimize a rare case:
1511 * Confirm(x, '=', Constlike) ==> Constlike
1513 if (is_irn_constlike(bound)) {
1514 DBG_OPT_CONFIRM(n, bound);
1518 return get_opt_remove_confirm() ? get_Confirm_value(n) : n;
1522 * Optimize CopyB(mem, x, x) into a Nop.
1524 static ir_node *equivalent_node_CopyB(ir_node *n) {
1525 ir_node *a = get_CopyB_dst(n);
1526 ir_node *b = get_CopyB_src(n);
1529 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1530 ir_node *mem = get_CopyB_mem(n);
1531 ir_node *blk = get_nodes_block(n);
1532 turn_into_tuple(n, pn_CopyB_max);
1533 set_Tuple_pred(n, pn_CopyB_M, mem);
1534 set_Tuple_pred(n, pn_CopyB_X_regular, new_r_Jmp(current_ir_graph, blk));
1535 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1536 set_Tuple_pred(n, pn_CopyB_M_except, new_Bad());
1539 } /* equivalent_node_CopyB */
1542 * Optimize Bounds(idx, idx, upper) into idx.
1544 static ir_node *equivalent_node_Bound(ir_node *n) {
1545 ir_node *idx = get_Bound_index(n);
1546 ir_node *lower = get_Bound_lower(n);
1549 /* By definition lower < upper, so if idx == lower -->
1550 lower <= idx && idx < upper */
1552 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1555 ir_node *pred = skip_Proj(idx);
1557 if (get_irn_op(pred) == op_Bound) {
1559 * idx was Bounds_check previously, it is still valid if
1560 * lower <= pred_lower && pred_upper <= upper.
1562 ir_node *upper = get_Bound_upper(n);
1563 if (get_Bound_lower(pred) == lower &&
1564 get_Bound_upper(pred) == upper) {
1566 * One could expect that we simply return the previous
1567 * Bound here. However, this would be wrong, as we could
1568 * add an exception Proj to a new location then.
1569 * So, we must turn in into a tuple.
1576 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1577 ir_node *mem = get_Bound_mem(n);
1578 ir_node *blk = get_nodes_block(n);
1579 turn_into_tuple(n, pn_Bound_max);
1580 set_Tuple_pred(n, pn_Bound_M, mem);
1581 set_Tuple_pred(n, pn_Bound_X_regular, new_r_Jmp(current_ir_graph, blk)); /* no exception */
1582 set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */
1583 set_Tuple_pred(n, pn_Bound_res, idx);
1586 } /* equivalent_node_Bound */
1589 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1590 * perform no actual computation, as, e.g., the Id nodes. It does not create
1591 * new nodes. It is therefore safe to free n if the node returned is not n.
1592 * If a node returns a Tuple we can not just skip it. If the size of the
1593 * in array fits, we transform n into a tuple (e.g., Div).
1595 ir_node *equivalent_node(ir_node *n) {
1596 if (n->op->ops.equivalent_node)
1597 return n->op->ops.equivalent_node(n);
1599 } /* equivalent_node */
1602 * Sets the default equivalent node operation for an ir_op_ops.
1604 * @param code the opcode for the default operation
1605 * @param ops the operations initialized
1610 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1614 ops->equivalent_node = equivalent_node_##a; \
1654 } /* firm_set_default_equivalent_node */
1657 * Do node specific optimizations of nodes predecessors.
1659 static void optimize_preds(ir_node *n) {
1660 switch (get_irn_opcode(n)) {
1662 case iro_Cmp: { /* We don't want Cast as input to Cmp. */
1663 ir_node *a = get_Cmp_left(n), *b = get_Cmp_right(n);
1665 if (get_irn_op(a) == op_Cast) {
1669 if (get_irn_op(b) == op_Cast) {
1671 set_Cmp_right(n, b);
1678 } /* optimize_preds */
1681 * Returns non-zero if a node is a Phi node
1682 * with all predecessors constant.
1684 static int is_const_Phi(ir_node *n) {
1689 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1690 if (! is_Const(get_irn_n(n, i)))
1693 } /* is_const_Phi */
1696 * Apply an evaluator on a binop with a constant operators (and one Phi).
1698 * @param phi the Phi node
1699 * @param other the other operand
1700 * @param eval an evaluator function
1701 * @param left if non-zero, other is the left operand, else the right
1703 * @return a new Phi node if the conversion was successful, NULL else
1705 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), int left) {
1711 int i, n = get_irn_arity(phi);
1713 NEW_ARR_A(void *, res, n);
1715 for (i = 0; i < n; ++i) {
1716 pred = get_irn_n(phi, i);
1717 tv = get_Const_tarval(pred);
1718 tv = eval(other, tv);
1720 if (tv == tarval_bad) {
1721 /* folding failed, bad */
1727 for (i = 0; i < n; ++i) {
1728 pred = get_irn_n(phi, i);
1729 tv = get_Const_tarval(pred);
1730 tv = eval(tv, other);
1732 if (tv == tarval_bad) {
1733 /* folding failed, bad */
1739 mode = get_irn_mode(phi);
1740 irg = current_ir_graph;
1741 for (i = 0; i < n; ++i) {
1742 pred = get_irn_n(phi, i);
1743 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1744 mode, res[i], get_Const_type(pred));
1746 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1747 } /* apply_binop_on_phi */
1750 * Apply an evaluator on a unop with a constant operator (a Phi).
1752 * @param phi the Phi node
1753 * @param eval an evaluator function
1755 * @return a new Phi node if the conversion was successful, NULL else
1757 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
1763 int i, n = get_irn_arity(phi);
1765 NEW_ARR_A(void *, res, n);
1766 for (i = 0; i < n; ++i) {
1767 pred = get_irn_n(phi, i);
1768 tv = get_Const_tarval(pred);
1771 if (tv == tarval_bad) {
1772 /* folding failed, bad */
1777 mode = get_irn_mode(phi);
1778 irg = current_ir_graph;
1779 for (i = 0; i < n; ++i) {
1780 pred = get_irn_n(phi, i);
1781 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1782 mode, res[i], get_Const_type(pred));
1784 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1785 } /* apply_unop_on_phi */
1788 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1789 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1790 * If possible, remove the Conv's.
1792 static ir_node *transform_node_AddSub(ir_node *n) {
1793 ir_mode *mode = get_irn_mode(n);
1795 if (mode_is_reference(mode)) {
1796 ir_node *left = get_binop_left(n);
1797 ir_node *right = get_binop_right(n);
1798 int ref_bits = get_mode_size_bits(mode);
1800 if (get_irn_op(left) == op_Conv) {
1801 ir_mode *mode = get_irn_mode(left);
1802 int bits = get_mode_size_bits(mode);
1804 if (ref_bits == bits &&
1805 mode_is_int(mode) &&
1806 get_mode_arithmetic(mode) == irma_twos_complement) {
1807 ir_node *pre = get_Conv_op(left);
1808 ir_mode *pre_mode = get_irn_mode(pre);
1810 if (mode_is_int(pre_mode) &&
1811 get_mode_size_bits(pre_mode) == bits &&
1812 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1813 /* ok, this conv just changes to sign, moreover the calculation
1814 * is done with same number of bits as our address mode, so
1815 * we can ignore the conv as address calculation can be viewed
1816 * as either signed or unsigned
1818 set_binop_left(n, pre);
1823 if (get_irn_op(right) == op_Conv) {
1824 ir_mode *mode = get_irn_mode(right);
1825 int bits = get_mode_size_bits(mode);
1827 if (ref_bits == bits &&
1828 mode_is_int(mode) &&
1829 get_mode_arithmetic(mode) == irma_twos_complement) {
1830 ir_node *pre = get_Conv_op(right);
1831 ir_mode *pre_mode = get_irn_mode(pre);
1833 if (mode_is_int(pre_mode) &&
1834 get_mode_size_bits(pre_mode) == bits &&
1835 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1836 /* ok, this conv just changes to sign, moreover the calculation
1837 * is done with same number of bits as our address mode, so
1838 * we can ignore the conv as address calculation can be viewed
1839 * as either signed or unsigned
1841 set_binop_right(n, pre);
1847 } /* transform_node_AddSub */
1849 #define HANDLE_BINOP_PHI(op,a,b,c) \
1851 if (is_Const(b) && is_const_Phi(a)) { \
1852 /* check for Op(Phi, Const) */ \
1853 c = apply_binop_on_phi(a, get_Const_tarval(b), op, 0); \
1855 else if (is_Const(a) && is_const_Phi(b)) { \
1856 /* check for Op(Const, Phi) */ \
1857 c = apply_binop_on_phi(b, get_Const_tarval(a), op, 1); \
1860 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1864 #define HANDLE_UNOP_PHI(op,a,c) \
1866 if (is_const_Phi(a)) { \
1867 /* check for Op(Phi) */ \
1868 c = apply_unop_on_phi(a, op); \
1871 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1877 * Do the AddSub optimization, then Transform
1878 * Constant folding on Phi
1879 * Add(a,a) -> Mul(a, 2)
1880 * Add(Mul(a, x), a) -> Mul(a, x+1)
1881 * if the mode is integer or float.
1882 * Transform Add(a,-b) into Sub(a,b).
1883 * Reassociation might fold this further.
1885 static ir_node *transform_node_Add(ir_node *n) {
1887 ir_node *a, *b, *c, *oldn = n;
1889 n = transform_node_AddSub(n);
1891 a = get_Add_left(n);
1892 b = get_Add_right(n);
1894 HANDLE_BINOP_PHI(tarval_add, a,b,c);
1896 mode = get_irn_mode(n);
1898 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1899 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1902 if (mode_is_num(mode)) {
1904 ir_node *block = get_irn_n(n, -1);
1907 get_irn_dbg_info(n),
1911 new_r_Const_long(current_ir_graph, block, mode, 2),
1913 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
1914 } else if (get_irn_op(a) == op_Minus) {
1916 get_irn_dbg_info(n),
1922 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1923 } else if (get_irn_op(b) == op_Minus) {
1925 get_irn_dbg_info(n),
1931 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1933 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1934 else if (!get_opt_reassociation() && get_irn_op(a) == op_Mul) {
1935 ir_node *ma = get_Mul_left(a);
1936 ir_node *mb = get_Mul_right(a);
1939 ir_node *blk = get_irn_n(n, -1);
1941 get_irn_dbg_info(n), current_ir_graph, blk,
1944 get_irn_dbg_info(n), current_ir_graph, blk,
1946 new_r_Const_long(current_ir_graph, blk, mode, 1),
1949 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1950 } else if (b == mb) {
1951 ir_node *blk = get_irn_n(n, -1);
1953 get_irn_dbg_info(n), current_ir_graph, blk,
1956 get_irn_dbg_info(n), current_ir_graph, blk,
1958 new_r_Const_long(current_ir_graph, blk, mode, 1),
1961 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1964 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1965 else if (!get_opt_reassociation() && get_irn_op(b) == op_Mul) {
1966 ir_node *ma = get_Mul_left(b);
1967 ir_node *mb = get_Mul_right(b);
1970 ir_node *blk = get_irn_n(n, -1);
1972 get_irn_dbg_info(n), current_ir_graph, blk,
1975 get_irn_dbg_info(n), current_ir_graph, blk,
1977 new_r_Const_long(current_ir_graph, blk, mode, 1),
1980 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1981 } else if (a == mb) {
1982 ir_node *blk = get_irn_n(n, -1);
1984 get_irn_dbg_info(n), current_ir_graph, blk,
1987 get_irn_dbg_info(n), current_ir_graph, blk,
1989 new_r_Const_long(current_ir_graph, blk, mode, 1),
1992 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1995 /* Here we rely on constants be on the RIGHT side */
1996 else if (is_Not(a) && classify_Const(b) == CNST_ONE) {
1998 ir_node *op = get_Not_op(a);
1999 ir_node *blk = get_irn_n(n, -1);
2000 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2001 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2005 } /* transform_node_Add */
2008 * Do the AddSub optimization, then Transform
2009 * Constant folding on Phi
2010 * Sub(0,a) -> Minus(a)
2011 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2012 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2014 static ir_node *transform_node_Sub(ir_node *n) {
2019 n = transform_node_AddSub(n);
2021 a = get_Sub_left(n);
2022 b = get_Sub_right(n);
2024 mode = get_irn_mode(n);
2027 HANDLE_BINOP_PHI(tarval_sub, a,b,c);
2029 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2030 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2034 if (mode_wrap_around(mode)) {
2035 ir_node *left = get_Add_left(a);
2036 ir_node *right = get_Add_right(a);
2038 /* FIXME: Does the Conv's word only for two complement or generally? */
2040 if (mode != get_irn_mode(right)) {
2041 /* This Sub is an effective Cast */
2042 right = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), right, mode);
2045 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2046 } else if (right == b) {
2047 if (mode != get_irn_mode(left)) {
2048 /* This Sub is an effective Cast */
2049 left = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), left, mode);
2052 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2055 } else if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2056 ir_mode *mode = get_irn_mode(a);
2058 if (mode == get_irn_mode(b)) {
2064 /* check if it's allowed to skip the conv */
2065 ma = get_irn_mode(a);
2066 mb = get_irn_mode(b);
2068 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2069 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2071 set_Sub_right(n, b);
2077 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2078 else if (mode_is_num(mode) && mode == get_irn_mode(a) && (classify_Const(a) == CNST_NULL)) {
2080 get_irn_dbg_info(n),
2085 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2087 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2088 else if (get_opt_reassociation() && get_irn_op(a) == op_Mul) {
2089 ir_node *ma = get_Mul_left(a);
2090 ir_node *mb = get_Mul_right(a);
2093 ir_node *blk = get_irn_n(n, -1);
2095 get_irn_dbg_info(n),
2096 current_ir_graph, blk,
2099 get_irn_dbg_info(n),
2100 current_ir_graph, blk,
2102 new_r_Const_long(current_ir_graph, blk, mode, 1),
2105 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2106 } else if (mb == b) {
2107 ir_node *blk = get_irn_n(n, -1);
2109 get_irn_dbg_info(n),
2110 current_ir_graph, blk,
2113 get_irn_dbg_info(n),
2114 current_ir_graph, blk,
2116 new_r_Const_long(current_ir_graph, blk, mode, 1),
2119 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2121 } else if (get_irn_op(a) == op_Sub) {
2122 ir_node *x = get_Sub_left(a);
2123 ir_node *y = get_Sub_right(a);
2124 ir_node *blk = get_irn_n(n, -1);
2125 ir_mode *m_b = get_irn_mode(b);
2126 ir_mode *m_y = get_irn_mode(y);
2129 /* Determine the right mode for the Add. */
2132 else if (mode_is_reference(m_b))
2134 else if (mode_is_reference(m_y))
2138 * Both modes are different but none is reference,
2139 * happens for instance in SubP(SubP(P, Iu), Is).
2140 * We have two possibilities here: Cast or ignore.
2141 * Currently we ignore this case.
2146 add = new_r_Add(current_ir_graph, blk, y, b, mode);
2149 set_Sub_right(n, add);
2150 DBG_OPT_ALGSIM0(n, n, FS_OPT_SUB_SUB_X_Y_Z);
2153 } /* transform_node_Sub */
2156 * Transform Mul(a,-1) into -a.
2157 * Do constant evaluation of Phi nodes.
2158 * Do architecture dependent optimizations on Mul nodes
2160 static ir_node *transform_node_Mul(ir_node *n) {
2161 ir_node *c, *oldn = n;
2162 ir_node *a = get_Mul_left(n);
2163 ir_node *b = get_Mul_right(n);
2166 HANDLE_BINOP_PHI(tarval_mul, a,b,c);
2168 mode = get_irn_mode(n);
2169 if (mode_is_signed(mode)) {
2172 if (value_of(a) == get_mode_minus_one(mode))
2174 else if (value_of(b) == get_mode_minus_one(mode))
2177 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2178 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2182 return arch_dep_replace_mul_with_shifts(n);
2183 } /* transform_node_Mul */
2186 * Transform a Div Node.
2188 static ir_node *transform_node_Div(ir_node *n) {
2189 tarval *tv = value_of(n);
2192 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2194 if (tv != tarval_bad) {
2195 value = new_Const(get_tarval_mode(tv), tv);
2197 DBG_OPT_CSTEVAL(n, value);
2198 } else /* Try architecture dependent optimization */
2199 value = arch_dep_replace_div_by_const(n);
2202 /* Turn Div into a tuple (mem, jmp, bad, value) */
2203 ir_node *mem = get_Div_mem(n);
2204 ir_node *blk = get_irn_n(n, -1);
2206 turn_into_tuple(n, pn_Div_max);
2207 set_Tuple_pred(n, pn_Div_M, mem);
2208 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2209 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2210 set_Tuple_pred(n, pn_Div_res, value);
2213 } /* transform_node_Div */
2216 * Transform a Mod node.
2218 static ir_node *transform_node_Mod(ir_node *n) {
2219 tarval *tv = value_of(n);
2222 /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */
2224 if (tv != tarval_bad) {
2225 value = new_Const(get_tarval_mode(tv), tv);
2227 DBG_OPT_CSTEVAL(n, value);
2228 } else /* Try architecture dependent optimization */
2229 value = arch_dep_replace_mod_by_const(n);
2232 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2233 ir_node *mem = get_Mod_mem(n);
2234 ir_node *blk = get_irn_n(n, -1);
2236 turn_into_tuple(n, pn_Mod_max);
2237 set_Tuple_pred(n, pn_Mod_M, mem);
2238 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2239 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2240 set_Tuple_pred(n, pn_Mod_res, value);
2243 } /* transform_node_Mod */
2246 * Transform a DivMod node.
2248 static ir_node *transform_node_DivMod(ir_node *n) {
2251 ir_node *a = get_DivMod_left(n);
2252 ir_node *b = get_DivMod_right(n);
2253 ir_mode *mode = get_irn_mode(a);
2254 tarval *ta = value_of(a);
2255 tarval *tb = value_of(b);
2257 if (!(mode_is_int(mode) && mode_is_int(get_irn_mode(b))))
2260 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2262 if (tb != tarval_bad) {
2263 if (tb == get_mode_one(get_tarval_mode(tb))) {
2264 b = new_Const (mode, get_mode_null(mode));
2267 DBG_OPT_CSTEVAL(n, b);
2268 } else if (ta != tarval_bad) {
2269 tarval *resa, *resb;
2270 resa = tarval_div (ta, tb);
2271 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2272 Jmp for X result!? */
2273 resb = tarval_mod (ta, tb);
2274 if (resb == tarval_bad) return n; /* Causes exception! */
2275 a = new_Const (mode, resa);
2276 b = new_Const (mode, resb);
2279 DBG_OPT_CSTEVAL(n, a);
2280 DBG_OPT_CSTEVAL(n, b);
2281 } else { /* Try architecture dependent optimization */
2282 arch_dep_replace_divmod_by_const(&a, &b, n);
2283 evaluated = a != NULL;
2285 } else if (ta == get_mode_null(mode)) {
2286 /* 0 / non-Const = 0 */
2291 if (evaluated) { /* replace by tuple */
2292 ir_node *mem = get_DivMod_mem(n);
2293 ir_node *blk = get_irn_n(n, -1);
2294 turn_into_tuple(n, pn_DivMod_max);
2295 set_Tuple_pred(n, pn_DivMod_M, mem);
2296 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
2297 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
2298 set_Tuple_pred(n, pn_DivMod_res_div, a);
2299 set_Tuple_pred(n, pn_DivMod_res_mod, b);
2303 } /* transform_node_DivMod */
2306 * Optimize Abs(x) into x if x is Confirmed >= 0
2307 * Optimize Abs(x) into -x if x is Confirmed <= 0
2309 static ir_node *transform_node_Abs(ir_node *n) {
2311 ir_node *a = get_Abs_op(n);
2312 value_classify_sign sign = classify_value_sign(a);
2314 if (sign == value_classified_negative) {
2315 ir_mode *mode = get_irn_mode(n);
2318 * We can replace the Abs by -x here.
2319 * We even could add a new Confirm here.
2321 * Note that -x would create a new node, so we could
2322 * not run it in the equivalent_node() context.
2324 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
2325 get_irn_n(n, -1), a, mode);
2327 DBG_OPT_CONFIRM(oldn, n);
2328 } else if (sign == value_classified_positive) {
2329 /* n is positive, Abs is not needed */
2332 DBG_OPT_CONFIRM(oldn, n);
2336 } /* transform_node_Abs */
2339 * Transform a Cond node.
2341 * Replace the Cond by a Jmp if it branches on a constant
2344 static ir_node *transform_node_Cond(ir_node *n) {
2347 ir_node *a = get_Cond_selector(n);
2348 tarval *ta = value_of(a);
2350 /* we need block info which is not available in floating irgs */
2351 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
2354 if ((ta != tarval_bad) &&
2355 (get_irn_mode(a) == mode_b) &&
2356 (get_opt_unreachable_code())) {
2357 /* It's a boolean Cond, branching on a boolean constant.
2358 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2359 jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n));
2360 turn_into_tuple(n, pn_Cond_max);
2361 if (ta == tarval_b_true) {
2362 set_Tuple_pred(n, pn_Cond_false, new_Bad());
2363 set_Tuple_pred(n, pn_Cond_true, jmp);
2365 set_Tuple_pred(n, pn_Cond_false, jmp);
2366 set_Tuple_pred(n, pn_Cond_true, new_Bad());
2368 /* We might generate an endless loop, so keep it alive. */
2369 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
2372 } /* transform_node_Cond */
2374 typedef ir_node* (*recursive_transform) (ir_node *n);
2377 * makes use of distributive laws for and, or, eor
2378 * and(a OP c, b OP c) -> and(a, b) OP c
2380 static ir_node *transform_bitwise_distributive(ir_node *n,
2381 recursive_transform trans_func)
2384 ir_node *a = get_binop_left(n);
2385 ir_node *b = get_binop_right(n);
2386 ir_op *op = get_irn_op(a);
2387 ir_op *op_root = get_irn_op(n);
2389 if(op != get_irn_op(b))
2392 if (op == op_Conv) {
2393 ir_node *a_op = get_Conv_op(a);
2394 ir_node *b_op = get_Conv_op(b);
2395 ir_mode *a_mode = get_irn_mode(a_op);
2396 ir_mode *b_mode = get_irn_mode(b_op);
2397 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
2398 ir_node *blk = get_irn_n(n, -1);
2401 set_binop_left(n, a_op);
2402 set_binop_right(n, b_op);
2403 set_irn_mode(n, a_mode);
2405 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
2407 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
2413 /* nothing to gain here */
2417 if (op == op_Shrs || op == op_Shr || op == op_Shl
2418 || op == op_And || op == op_Or || op == op_Eor) {
2419 ir_node *a_left = get_binop_left(a);
2420 ir_node *a_right = get_binop_right(a);
2421 ir_node *b_left = get_binop_left(b);
2422 ir_node *b_right = get_binop_right(b);
2426 if (is_op_commutative(op)) {
2427 if (a_left == b_left) {
2431 } else if(a_left == b_right) {
2435 } else if(a_right == b_left) {
2441 if(a_right == b_right) {
2448 /* (a sop c) & (b sop c) => (a & b) sop c */
2449 ir_node *blk = get_irn_n(n, -1);
2451 ir_node *new_n = exact_copy(n);
2452 set_binop_left(new_n, op1);
2453 set_binop_right(new_n, op2);
2454 new_n = trans_func(new_n);
2456 if(op_root == op_Eor && op == op_Or) {
2457 dbg_info *dbgi = get_irn_dbg_info(n);
2458 ir_graph *irg = current_ir_graph;
2459 ir_mode *mode = get_irn_mode(c);
2461 c = new_rd_Not(dbgi, irg, blk, c, mode);
2462 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
2465 set_irn_n(n, -1, blk);
2466 set_binop_left(n, new_n);
2467 set_binop_right(n, c);
2471 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
2482 static ir_node *transform_node_And(ir_node *n) {
2484 ir_node *a = get_And_left(n);
2485 ir_node *b = get_And_right(n);
2487 HANDLE_BINOP_PHI(tarval_and, a,b,c);
2489 n = transform_bitwise_distributive(n, transform_node_And);
2492 } /* transform_node_And */
2497 static ir_node *transform_node_Eor(ir_node *n) {
2498 ir_node *c, *oldn = n;
2499 ir_node *a = get_Eor_left(n);
2500 ir_node *b = get_Eor_right(n);
2501 ir_mode *mode = get_irn_mode(n);
2503 HANDLE_BINOP_PHI(tarval_eor, a,b,c);
2507 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
2508 mode, get_mode_null(mode));
2509 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
2510 } else if ((mode == mode_b)
2511 && (get_irn_op(a) == op_Proj)
2512 && (get_irn_mode(a) == mode_b)
2513 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)
2514 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2515 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
2516 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2517 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
2519 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
2520 } else if ((mode == mode_b)
2521 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) {
2522 /* The Eor is a Not. Replace it by a Not. */
2523 /* ????!!!Extend to bitfield 1111111. */
2524 n = new_r_Not(current_ir_graph, get_irn_n(n, -1), a, mode_b);
2526 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2528 n = transform_bitwise_distributive(n, transform_node_Eor);
2532 } /* transform_node_Eor */
2537 static ir_node *transform_node_Not(ir_node *n) {
2538 ir_node *c, *oldn = n;
2539 ir_node *a = get_Not_op(n);
2540 ir_op *op_a = get_irn_op(a);
2542 HANDLE_UNOP_PHI(tarval_not,a,c);
2544 /* check for a boolean Not */
2545 if ( (get_irn_mode(n) == mode_b)
2546 && (op_a == op_Proj)
2547 && (get_irn_mode(a) == mode_b)
2548 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2549 /* We negate a Cmp. The Cmp has the negated result anyways! */
2550 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2551 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
2552 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
2555 if (op_a == op_Sub && classify_Const(get_Sub_right(a)) == CNST_ONE) {
2557 ir_node *op = get_Sub_left(a);
2558 ir_node *blk = get_irn_n(n, -1);
2559 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
2560 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
2563 } /* transform_node_Not */
2566 * Transform a Minus.
2570 static ir_node *transform_node_Minus(ir_node *n) {
2571 ir_node *c, *oldn = n;
2572 ir_node *a = get_Minus_op(n);
2574 HANDLE_UNOP_PHI(tarval_neg,a,c);
2578 ir_node *op = get_Not_op(a);
2579 ir_mode *mode = get_irn_mode(op);
2580 tarval *tv = get_mode_one(mode);
2581 ir_node *blk = get_irn_n(n, -1);
2582 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2583 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
2584 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
2588 } /* transform_node_Minus */
2591 * Transform a Cast_type(Const) into a new Const_type
2593 static ir_node *transform_node_Cast(ir_node *n) {
2595 ir_node *pred = get_Cast_op(n);
2596 ir_type *tp = get_irn_type(n);
2598 if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) {
2599 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
2600 get_Const_tarval(pred), tp);
2601 DBG_OPT_CSTEVAL(oldn, n);
2602 } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) {
2603 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred),
2604 get_SymConst_kind(pred), tp);
2605 DBG_OPT_CSTEVAL(oldn, n);
2609 } /* transform_node_Cast */
2612 * Transform a Proj(Div) with a non-zero value.
2613 * Removes the exceptions and routes the memory to the NoMem node.
2615 static ir_node *transform_node_Proj_Div(ir_node *proj) {
2616 ir_node *div = get_Proj_pred(proj);
2617 ir_node *b = get_Div_right(div);
2618 ir_node *confirm, *res, *new_mem;
2621 if (value_not_zero(b, &confirm)) {
2622 /* div(x, y) && y != 0 */
2623 proj_nr = get_Proj_proj(proj);
2625 case pn_Div_X_regular:
2626 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
2628 case pn_Div_X_except:
2629 /* we found an exception handler, remove it */
2630 DBG_OPT_EXC_REM(proj);
2634 res = get_Div_mem(div);
2635 new_mem = get_irg_no_mem(current_ir_graph);
2638 /* This node can only float up to the Confirm block */
2639 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2641 set_irn_pinned(div, op_pin_state_floats);
2642 /* this is a Div without exception, we can remove the memory edge */
2643 set_Div_mem(div, new_mem);
2648 } /* transform_node_Proj_Div */
2651 * Transform a Proj(Mod) with a non-zero value.
2652 * Removes the exceptions and routes the memory to the NoMem node.
2654 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
2655 ir_node *mod = get_Proj_pred(proj);
2656 ir_node *b = get_Mod_right(mod);
2657 ir_node *confirm, *res, *new_mem;
2660 if (value_not_zero(b, &confirm)) {
2661 /* mod(x, y) && y != 0 */
2662 proj_nr = get_Proj_proj(proj);
2666 case pn_Mod_X_regular:
2667 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
2669 case pn_Mod_X_except:
2670 /* we found an exception handler, remove it */
2671 DBG_OPT_EXC_REM(proj);
2675 res = get_Mod_mem(mod);
2676 new_mem = get_irg_no_mem(current_ir_graph);
2679 /* This node can only float up to the Confirm block */
2680 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2682 set_irn_pinned(mod, op_pin_state_floats);
2683 /* this is a Mod without exception, we can remove the memory edge */
2684 set_Mod_mem(mod, get_irg_no_mem(current_ir_graph));
2687 if (get_Mod_left(mod) == b) {
2688 /* a % a = 0 if a != 0 */
2689 ir_mode *mode = get_irn_mode(proj);
2690 ir_node *res = new_Const(mode, get_mode_null(mode));
2692 DBG_OPT_CSTEVAL(mod, res);
2698 } /* transform_node_Proj_Mod */
2701 * Transform a Proj(DivMod) with a non-zero value.
2702 * Removes the exceptions and routes the memory to the NoMem node.
2704 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
2705 ir_node *divmod = get_Proj_pred(proj);
2706 ir_node *b = get_DivMod_right(divmod);
2707 ir_node *confirm, *res, *new_mem;
2710 if (value_not_zero(b, &confirm)) {
2711 /* DivMod(x, y) && y != 0 */
2712 proj_nr = get_Proj_proj(proj);
2716 case pn_DivMod_X_regular:
2717 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
2719 case pn_DivMod_X_except:
2720 /* we found an exception handler, remove it */
2721 DBG_OPT_EXC_REM(proj);
2725 res = get_DivMod_mem(divmod);
2726 new_mem = get_irg_no_mem(current_ir_graph);
2729 /* This node can only float up to the Confirm block */
2730 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2732 set_irn_pinned(divmod, op_pin_state_floats);
2733 /* this is a DivMod without exception, we can remove the memory edge */
2734 set_DivMod_mem(divmod, get_irg_no_mem(current_ir_graph));
2737 case pn_DivMod_res_mod:
2738 if (get_DivMod_left(divmod) == b) {
2739 /* a % a = 0 if a != 0 */
2740 ir_mode *mode = get_irn_mode(proj);
2741 ir_node *res = new_Const(mode, get_mode_null(mode));
2743 DBG_OPT_CSTEVAL(divmod, res);
2749 } /* transform_node_Proj_DivMod */
2752 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
2754 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
2755 if (get_opt_unreachable_code()) {
2756 ir_node *n = get_Proj_pred(proj);
2757 ir_node *b = get_Cond_selector(n);
2759 if (mode_is_int(get_irn_mode(b))) {
2760 tarval *tb = value_of(b);
2762 if (tb != tarval_bad) {
2763 /* we have a constant switch */
2764 long num = get_Proj_proj(proj);
2766 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
2767 if (get_tarval_long(tb) == num) {
2768 /* Do NOT create a jump here, or we will have 2 control flow ops
2769 * in a block. This case is optimized away in optimize_cf(). */
2772 /* this case will NEVER be taken, kill it */
2780 } /* transform_node_Proj_Cond */
2783 * Normalizes and optimizes Cmp nodes.
2785 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
2786 if (get_opt_reassociation()) {
2787 ir_node *n = get_Proj_pred(proj);
2788 ir_node *left = get_Cmp_left(n);
2789 ir_node *right = get_Cmp_right(n);
2793 ir_mode *mode = NULL;
2794 long proj_nr = get_Proj_proj(proj);
2797 * First step: normalize the compare op
2798 * by placing the constant on the right site
2799 * or moving the lower address node to the left.
2800 * We ignore the case that both are constants
2801 * this case should be optimized away.
2803 if (get_irn_op(right) == op_Const) {
2805 } else if (get_irn_op(left) == op_Const) {
2810 proj_nr = get_inversed_pnc(proj_nr);
2812 } else if (get_irn_idx(left) > get_irn_idx(right)) {
2818 proj_nr = get_inversed_pnc(proj_nr);
2823 * Second step: Try to reduce the magnitude
2824 * of a constant. This may help to generate better code
2825 * later and may help to normalize more compares.
2826 * Of course this is only possible for integer values.
2829 mode = get_irn_mode(c);
2830 tv = get_Const_tarval(c);
2832 if (tv != tarval_bad) {
2833 /* the following optimization is possible on modes without Overflow
2834 * on Unary Minus or on == and !=:
2835 * -a CMP c ==> a swap(CMP) -c
2837 * Beware: for two-complement Overflow may occur, so only == and != can
2838 * be optimized, see this:
2839 * -MININT < 0 =/=> MININT > 0 !!!
2841 if (get_opt_constant_folding() && get_irn_op(left) == op_Minus &&
2842 (!mode_overflow_on_unary_Minus(mode) ||
2843 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
2844 left = get_Minus_op(left);
2845 tv = tarval_neg(tv);
2847 if (tv != tarval_bad) {
2848 proj_nr = get_inversed_pnc(proj_nr);
2853 /* for integer modes, we have more */
2854 if (mode_is_int(mode)) {
2855 /* Ne includes Unordered which is not possible on integers.
2856 * However, frontends often use this wrong, so fix it here */
2857 if (proj_nr & pn_Cmp_Uo) {
2858 proj_nr &= ~pn_Cmp_Uo;
2859 set_Proj_proj(proj, proj_nr);
2862 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
2863 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
2864 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
2865 tv = tarval_sub(tv, get_mode_one(mode));
2867 if (tv != tarval_bad) {
2868 proj_nr ^= pn_Cmp_Eq;
2872 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
2873 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
2874 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
2875 tv = tarval_add(tv, get_mode_one(mode));
2877 if (tv != tarval_bad) {
2878 proj_nr ^= pn_Cmp_Eq;
2883 /* the following reassociations work only for == and != */
2884 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
2886 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
2887 if (classify_tarval(tv) == TV_CLASSIFY_NULL && get_irn_op(left) == op_Sub) {
2888 right = get_Sub_right(left);
2889 left = get_Sub_left(left);
2891 tv = value_of(right);
2892 if (tv != tarval_bad) {
2897 if (tv != tarval_bad) {
2898 ir_op *op = get_irn_op(left);
2900 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
2902 ir_node *c1 = get_Sub_right(left);
2903 tarval *tv2 = value_of(c1);
2905 if (tv2 != tarval_bad) {
2906 tv2 = tarval_add(tv, value_of(c1));
2908 if (tv2 != tarval_bad) {
2909 left = get_Sub_left(left);
2915 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
2916 else if (op == op_Add) {
2917 ir_node *a_l = get_Add_left(left);
2918 ir_node *a_r = get_Add_right(left);
2922 if (get_irn_op(a_l) == op_Const) {
2924 tv2 = value_of(a_l);
2927 tv2 = value_of(a_r);
2930 if (tv2 != tarval_bad) {
2931 tv2 = tarval_sub(tv, tv2);
2933 if (tv2 != tarval_bad) {
2940 /* -a == c ==> a == -c, -a != c ==> a != -c */
2941 else if (op == op_Minus) {
2942 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
2944 if (tv2 != tarval_bad) {
2945 left = get_Minus_op(left);
2952 /* the following reassociations work only for <= */
2953 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2954 if (tv != tarval_bad) {
2955 ir_op *op = get_irn_op(left);
2957 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
2965 * optimization for AND:
2967 * And(x, C) == C ==> And(x, C) != 0
2968 * And(x, C) != C ==> And(X, C) == 0
2970 * if C is a single Bit constant.
2972 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) &&
2973 (get_irn_op(left) == op_And)) {
2974 if (is_single_bit_tarval(tv)) {
2975 /* check for Constant's match. We have check hare the tarvals,
2976 because our const might be changed */
2977 ir_node *la = get_And_left(left);
2978 ir_node *ra = get_And_right(left);
2979 if ((is_Const(la) && get_Const_tarval(la) == tv) ||
2980 (is_Const(ra) && get_Const_tarval(ra) == tv)) {
2981 /* fine: do the transformation */
2982 tv = get_mode_null(get_tarval_mode(tv));
2983 proj_nr ^= pn_Cmp_Leg;
2988 } /* tarval != bad */
2992 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
2994 if (changed & 2) /* need a new Const */
2995 right = new_Const(mode, tv);
2997 /* create a new compare */
2998 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block,
3001 set_Proj_pred(proj, n);
3002 set_Proj_proj(proj, proj_nr);
3006 } /* transform_node_Proj_Cmp */
3009 * Does all optimizations on nodes that must be done on it's Proj's
3010 * because of creating new nodes.
3012 static ir_node *transform_node_Proj(ir_node *proj) {
3013 ir_node *n = get_Proj_pred(proj);
3015 switch (get_irn_opcode(n)) {
3017 return transform_node_Proj_Div(proj);
3020 return transform_node_Proj_Mod(proj);
3023 return transform_node_Proj_DivMod(proj);
3026 return transform_node_Proj_Cond(proj);
3029 return transform_node_Proj_Cmp(proj);
3032 /* should not happen, but if it does will be optimized away */
3033 return equivalent_node_Proj(proj);
3039 } /* transform_node_Proj */
3042 * Move Confirms down through Phi nodes.
3044 static ir_node *transform_node_Phi(ir_node *phi) {
3046 ir_mode *mode = get_irn_mode(phi);
3048 if (mode_is_reference(mode)) {
3049 n = get_irn_arity(phi);
3051 /* Beware of Phi0 */
3053 ir_node *pred = get_irn_n(phi, 0);
3054 ir_node *bound, *new_Phi, *block, **in;
3057 if (! is_Confirm(pred))
3060 bound = get_Confirm_bound(pred);
3061 pnc = get_Confirm_cmp(pred);
3063 NEW_ARR_A(ir_node *, in, n);
3064 in[0] = get_Confirm_value(pred);
3066 for (i = 1; i < n; ++i) {
3067 pred = get_irn_n(phi, i);
3069 if (! is_Confirm(pred) ||
3070 get_Confirm_bound(pred) != bound ||
3071 get_Confirm_cmp(pred) != pnc)
3073 in[i] = get_Confirm_value(pred);
3075 /* move the Confirm nodes "behind" the Phi */
3076 block = get_irn_n(phi, -1);
3077 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
3078 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
3082 } /* transform_node_Phi */
3085 * Returns the operands of a commutative bin-op, if one operand is
3086 * a const, it is returned as the second one.
3088 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
3089 ir_node *op_a = get_binop_left(binop);
3090 ir_node *op_b = get_binop_right(binop);
3092 assert(is_op_commutative(get_irn_op(binop)));
3094 if (get_irn_op(op_a) == op_Const) {
3101 } /* get_comm_Binop_Ops */
3104 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
3105 * Such pattern may arise in bitfield stores.
3107 * value c4 value c4 & c2
3108 * AND c3 AND c1 | c3
3113 static ir_node *transform_node_Or_bf_store(ir_node *or) {
3116 ir_node *and_l, *c3;
3117 ir_node *value, *c4;
3118 ir_node *new_and, *new_const, *block;
3119 ir_mode *mode = get_irn_mode(or);
3121 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
3123 get_comm_Binop_Ops(or, &and, &c1);
3124 if ((get_irn_op(c1) != op_Const) || (get_irn_op(and) != op_And))
3127 get_comm_Binop_Ops(and, &or_l, &c2);
3128 if ((get_irn_op(c2) != op_Const) || (get_irn_op(or_l) != op_Or))
3131 get_comm_Binop_Ops(or_l, &and_l, &c3);
3132 if ((get_irn_op(c3) != op_Const) || (get_irn_op(and_l) != op_And))
3135 get_comm_Binop_Ops(and_l, &value, &c4);
3136 if (get_irn_op(c4) != op_Const)
3139 /* ok, found the pattern, check for conditions */
3140 assert(mode == get_irn_mode(and));
3141 assert(mode == get_irn_mode(or_l));
3142 assert(mode == get_irn_mode(and_l));
3144 tv1 = get_Const_tarval(c1);
3145 tv2 = get_Const_tarval(c2);
3146 tv3 = get_Const_tarval(c3);
3147 tv4 = get_Const_tarval(c4);
3149 tv = tarval_or(tv4, tv2);
3150 if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) {
3151 /* have at least one 0 at the same bit position */
3155 n_tv4 = tarval_not(tv4);
3156 if (tv3 != tarval_and(tv3, n_tv4)) {
3157 /* bit in the or_mask is outside the and_mask */
3161 n_tv2 = tarval_not(tv2);
3162 if (tv1 != tarval_and(tv1, n_tv2)) {
3163 /* bit in the or_mask is outside the and_mask */
3167 /* ok, all conditions met */
3168 block = get_irn_n(or, -1);
3170 new_and = new_r_And(current_ir_graph, block,
3171 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
3173 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
3175 set_Or_left(or, new_and);
3176 set_Or_right(or, new_const);
3178 /* check for more */
3179 return transform_node_Or_bf_store(or);
3180 } /* transform_node_Or_bf_store */
3183 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
3185 static ir_node *transform_node_Or_Rot(ir_node *or) {
3186 ir_mode *mode = get_irn_mode(or);
3187 ir_node *shl, *shr, *block;
3188 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
3191 if (! mode_is_int(mode))
3194 shl = get_binop_left(or);
3195 shr = get_binop_right(or);
3197 if (get_irn_op(shl) == op_Shr) {
3198 if (get_irn_op(shr) != op_Shl)
3204 } else if (get_irn_op(shl) != op_Shl) {
3206 } else if (get_irn_op(shr) != op_Shr) {
3209 x = get_Shl_left(shl);
3210 if (x != get_Shr_left(shr))
3213 c1 = get_Shl_right(shl);
3214 c2 = get_Shr_right(shr);
3215 if (get_irn_op(c1) == op_Const && get_irn_op(c2) == op_Const) {
3216 tv1 = get_Const_tarval(c1);
3217 if (! tarval_is_long(tv1))
3220 tv2 = get_Const_tarval(c2);
3221 if (! tarval_is_long(tv2))
3224 if (get_tarval_long(tv1) + get_tarval_long(tv2)
3225 != get_mode_size_bits(mode))
3228 /* yet, condition met */
3229 block = get_irn_n(or, -1);
3231 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
3233 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
3235 } else if (get_irn_op(c1) == op_Sub) {
3239 if (get_Sub_right(sub) != v)
3242 c1 = get_Sub_left(sub);
3243 if (get_irn_op(c1) != op_Const)
3246 tv1 = get_Const_tarval(c1);
3247 if (! tarval_is_long(tv1))
3250 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
3253 /* yet, condition met */
3254 block = get_nodes_block(or);
3256 /* a Rot right is not supported, so use a rot left */
3257 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
3259 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
3261 } else if (get_irn_op(c2) == op_Sub) {
3265 c1 = get_Sub_left(sub);
3266 if (get_irn_op(c1) != op_Const)
3269 tv1 = get_Const_tarval(c1);
3270 if (! tarval_is_long(tv1))
3273 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
3276 /* yet, condition met */
3277 block = get_irn_n(or, -1);
3280 n = new_r_Rot(current_ir_graph, block, x, v, mode);
3282 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
3287 } /* transform_node_Or_Rot */
3292 static ir_node *transform_node_Or(ir_node *n) {
3293 ir_node *c, *oldn = n;
3294 ir_node *a = get_Or_left(n);
3295 ir_node *b = get_Or_right(n);
3297 HANDLE_BINOP_PHI(tarval_or, a,b,c);
3299 n = transform_node_Or_bf_store(n);
3300 n = transform_node_Or_Rot(n);
3304 n = transform_bitwise_distributive(n, transform_node_Or);
3307 } /* transform_node_Or */
3311 static ir_node *transform_node(ir_node *n);
3314 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl.
3316 * Should be moved to reassociation?
3318 static ir_node *transform_node_shift(ir_node *n) {
3319 ir_node *left, *right;
3320 tarval *tv1, *tv2, *res;
3322 int modulo_shf, flag;
3324 left = get_binop_left(n);
3326 /* different operations */
3327 if (get_irn_op(left) != get_irn_op(n))
3330 right = get_binop_right(n);
3331 tv1 = value_of(right);
3332 if (tv1 == tarval_bad)
3335 tv2 = value_of(get_binop_right(left));
3336 if (tv2 == tarval_bad)
3339 res = tarval_add(tv1, tv2);
3341 /* beware: a simple replacement works only, if res < modulo shift */
3342 mode = get_irn_mode(n);
3346 modulo_shf = get_mode_modulo_shift(mode);
3347 if (modulo_shf > 0) {
3348 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
3350 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
3356 /* ok, we can replace it */
3357 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
3359 in[0] = get_binop_left(left);
3360 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
3362 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
3364 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
3366 return transform_node(irn);
3369 } /* transform_node_shift */
3374 static ir_node *transform_node_Shr(ir_node *n) {
3375 ir_node *c, *oldn = n;
3376 ir_node *a = get_Shr_left(n);
3377 ir_node *b = get_Shr_right(n);
3379 HANDLE_BINOP_PHI(tarval_shr, a, b, c);
3380 return transform_node_shift(n);
3381 } /* transform_node_Shr */
3386 static ir_node *transform_node_Shrs(ir_node *n) {
3387 ir_node *c, *oldn = n;
3388 ir_node *a = get_Shrs_left(n);
3389 ir_node *b = get_Shrs_right(n);
3391 HANDLE_BINOP_PHI(tarval_shrs, a, b, c);
3392 return transform_node_shift(n);
3393 } /* transform_node_Shrs */
3398 static ir_node *transform_node_Shl(ir_node *n) {
3399 ir_node *c, *oldn = n;
3400 ir_node *a = get_Shl_left(n);
3401 ir_node *b = get_Shl_right(n);
3403 HANDLE_BINOP_PHI(tarval_shl, a, b, c);
3404 return transform_node_shift(n);
3405 } /* transform_node_Shl */
3408 * Remove dead blocks and nodes in dead blocks
3409 * in keep alive list. We do not generate a new End node.
3411 static ir_node *transform_node_End(ir_node *n) {
3412 int i, j, n_keepalives = get_End_n_keepalives(n);
3415 NEW_ARR_A(ir_node *, in, n_keepalives);
3417 for (i = j = 0; i < n_keepalives; ++i) {
3418 ir_node *ka = get_End_keepalive(n, i);
3420 if (! is_Block_dead(ka)) {
3424 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
3427 /* FIXME: beabi need to keep a Proj(M) */
3428 if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka))
3431 if (j != n_keepalives)
3432 set_End_keepalives(n, j, in);
3434 } /* transform_node_End */
3437 * Optimize a Mux into some simpler cases.
3439 static ir_node *transform_node_Mux(ir_node *n) {
3440 ir_node *oldn = n, *sel = get_Mux_sel(n);
3441 ir_mode *mode = get_irn_mode(n);
3443 if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(mode)) {
3444 ir_node *cmp = get_Proj_pred(sel);
3445 long proj_nr = get_Proj_proj(sel);
3446 ir_node *f = get_Mux_false(n);
3447 ir_node *t = get_Mux_true(n);
3449 if (get_irn_op(cmp) == op_Cmp && classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
3450 ir_node *block = get_irn_n(n, -1);
3453 * Note: normalization puts the constant on the right site,
3454 * so we check only one case.
3456 * Note further that these optimization work even for floating point
3457 * with NaN's because -NaN == NaN.
3458 * However, if +0 and -0 is handled differently, we cannot use the first one.
3460 if (get_irn_op(f) == op_Minus &&
3461 get_Minus_op(f) == t &&
3462 get_Cmp_left(cmp) == t) {
3464 if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3465 /* Mux(a >=/> 0, -a, a) ==> Abs(a) */
3466 n = new_rd_Abs(get_irn_dbg_info(n),
3470 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3472 } else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3473 /* Mux(a <=/< 0, -a, a) ==> Minus(Abs(a)) */
3474 n = new_rd_Abs(get_irn_dbg_info(n),
3478 n = new_rd_Minus(get_irn_dbg_info(n),
3483 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3486 } else if (get_irn_op(t) == op_Minus &&
3487 get_Minus_op(t) == f &&
3488 get_Cmp_left(cmp) == f) {
3490 if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3491 /* Mux(a <=/< 0, a, -a) ==> Abs(a) */
3492 n = new_rd_Abs(get_irn_dbg_info(n),
3496 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3498 } else if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3499 /* Mux(a >=/> 0, a, -a) ==> Minus(Abs(a)) */
3500 n = new_rd_Abs(get_irn_dbg_info(n),
3504 n = new_rd_Minus(get_irn_dbg_info(n),
3509 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3514 if (mode_is_int(mode) && mode_is_signed(mode) &&
3515 get_mode_arithmetic(mode) == irma_twos_complement) {
3516 ir_node *x = get_Cmp_left(cmp);
3518 /* the following optimization works only with signed integer two-complement mode */
3520 if (mode == get_irn_mode(x)) {
3522 * FIXME: this restriction is two rigid, as it would still
3523 * work if mode(x) = Hs and mode == Is, but at least it removes
3526 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Le) &&
3527 classify_Const(t) == CNST_ALL_ONE &&
3528 classify_Const(f) == CNST_NULL) {
3530 * Mux(x:T </<= 0, 0, -1) -> Shrs(x, sizeof_bits(T) - 1)
3534 n = new_rd_Shrs(get_irn_dbg_info(n),
3535 current_ir_graph, block, x,
3536 new_r_Const_long(current_ir_graph, block, mode_Iu,
3537 get_mode_size_bits(mode) - 1),
3539 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3541 } else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) &&
3542 classify_Const(t) == CNST_ONE &&
3543 classify_Const(f) == CNST_NULL) {
3545 * Mux(x:T >/>= 0, 0, 1) -> Shr(-x, sizeof_bits(T) - 1)
3549 n = new_rd_Shr(get_irn_dbg_info(n),
3550 current_ir_graph, block,
3551 new_r_Minus(current_ir_graph, block, x, mode),
3552 new_r_Const_long(current_ir_graph, block, mode_Iu,
3553 get_mode_size_bits(mode) - 1),
3555 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3562 return arch_transform_node_Mux(n);
3563 } /* transform_node_Mux */
3566 * Optimize a Psi into some simpler cases.
3568 static ir_node *transform_node_Psi(ir_node *n) {
3570 return transform_node_Mux(n);
3573 } /* transform_node_Psi */
3576 * Tries several [inplace] [optimizing] transformations and returns an
3577 * equivalent node. The difference to equivalent_node() is that these
3578 * transformations _do_ generate new nodes, and thus the old node must
3579 * not be freed even if the equivalent node isn't the old one.
3581 static ir_node *transform_node(ir_node *n) {
3585 * Transform_node is the only "optimizing transformation" that might
3586 * return a node with a different opcode. We iterate HERE until fixpoint
3587 * to get the final result.
3591 if (n->op->ops.transform_node)
3592 n = n->op->ops.transform_node(n);
3593 } while (oldn != n);
3596 } /* transform_node */
3599 * Sets the default transform node operation for an ir_op_ops.
3601 * @param code the opcode for the default operation
3602 * @param ops the operations initialized
3607 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
3611 ops->transform_node = transform_node_##a; \
3644 } /* firm_set_default_transform_node */
3647 /* **************** Common Subexpression Elimination **************** */
3649 /** The size of the hash table used, should estimate the number of nodes
3651 #define N_IR_NODES 512
3653 /** Compares the attributes of two Const nodes. */
3654 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
3655 return (get_Const_tarval(a) != get_Const_tarval(b))
3656 || (get_Const_type(a) != get_Const_type(b));
3657 } /* node_cmp_attr_Const */
3659 /** Compares the attributes of two Proj nodes. */
3660 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
3661 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
3662 } /* node_cmp_attr_Proj */
3664 /** Compares the attributes of two Filter nodes. */
3665 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
3666 return get_Filter_proj(a) != get_Filter_proj(b);
3667 } /* node_cmp_attr_Filter */
3669 /** Compares the attributes of two Alloc nodes. */
3670 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
3671 const alloc_attr *pa = get_irn_alloc_attr(a);
3672 const alloc_attr *pb = get_irn_alloc_attr(b);
3673 return (pa->where != pb->where) || (pa->type != pb->type);
3674 } /* node_cmp_attr_Alloc */
3676 /** Compares the attributes of two Free nodes. */
3677 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
3678 const free_attr *pa = get_irn_free_attr(a);
3679 const free_attr *pb = get_irn_free_attr(b);
3680 return (pa->where != pb->where) || (pa->type != pb->type);
3681 } /* node_cmp_attr_Free */
3683 /** Compares the attributes of two SymConst nodes. */
3684 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
3685 const symconst_attr *pa = get_irn_symconst_attr(a);
3686 const symconst_attr *pb = get_irn_symconst_attr(b);
3687 return (pa->num != pb->num)
3688 || (pa->sym.type_p != pb->sym.type_p)
3689 || (pa->tp != pb->tp);
3690 } /* node_cmp_attr_SymConst */
3692 /** Compares the attributes of two Call nodes. */
3693 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
3694 return (get_irn_call_attr(a) != get_irn_call_attr(b));
3695 } /* node_cmp_attr_Call */
3697 /** Compares the attributes of two Sel nodes. */
3698 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
3699 const ir_entity *a_ent = get_Sel_entity(a);
3700 const ir_entity *b_ent = get_Sel_entity(b);
3702 (a_ent->kind != b_ent->kind) ||
3703 (a_ent->name != b_ent->name) ||
3704 (a_ent->owner != b_ent->owner) ||
3705 (a_ent->ld_name != b_ent->ld_name) ||
3706 (a_ent->type != b_ent->type);
3707 } /* node_cmp_attr_Sel */
3709 /** Compares the attributes of two Phi nodes. */
3710 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
3711 /* we can only enter this function if both nodes have the same number of inputs,
3712 hence it is enough to check if one of them is a Phi0 */
3714 /* check the Phi0 attribute */
3715 return get_irn_phi0_attr(a) != get_irn_phi0_attr(b);
3718 } /* node_cmp_attr_Phi */
3720 /** Compares the attributes of two Conv nodes. */
3721 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
3722 return get_Conv_strict(a) != get_Conv_strict(b);
3723 } /* node_cmp_attr_Conv */
3725 /** Compares the attributes of two Cast nodes. */
3726 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
3727 return get_Cast_type(a) != get_Cast_type(b);
3728 } /* node_cmp_attr_Cast */
3730 /** Compares the attributes of two Load nodes. */
3731 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
3732 if (get_Load_volatility(a) == volatility_is_volatile ||
3733 get_Load_volatility(b) == volatility_is_volatile)
3734 /* NEVER do CSE on volatile Loads */
3737 return get_Load_mode(a) != get_Load_mode(b);
3738 } /* node_cmp_attr_Load */
3740 /** Compares the attributes of two Store nodes. */
3741 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
3742 /* NEVER do CSE on volatile Stores */
3743 return (get_Store_volatility(a) == volatility_is_volatile ||
3744 get_Store_volatility(b) == volatility_is_volatile);
3745 } /* node_cmp_attr_Store */
3747 /** Compares the attributes of two Confirm nodes. */
3748 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
3749 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
3750 } /* node_cmp_attr_Confirm */
3752 /** Compares the attributes of two ASM nodes. */
3753 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
3755 const ir_asm_constraint *ca;
3756 const ir_asm_constraint *cb;
3759 if (get_ASM_text(a) != get_ASM_text(b))
3762 /* Should we really check the constraints here? Should be better, but is strange. */
3763 n = get_ASM_n_input_constraints(a);
3764 if (n != get_ASM_n_input_constraints(b))
3767 ca = get_ASM_input_constraints(a);
3768 cb = get_ASM_input_constraints(b);
3769 for (i = 0; i < n; ++i) {
3770 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
3774 n = get_ASM_n_output_constraints(a);
3775 if (n != get_ASM_n_output_constraints(b))
3778 ca = get_ASM_output_constraints(a);
3779 cb = get_ASM_output_constraints(b);
3780 for (i = 0; i < n; ++i) {
3781 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
3785 n = get_ASM_n_clobbers(a);
3786 if (n != get_ASM_n_clobbers(b))
3789 cla = get_ASM_clobbers(a);
3790 clb = get_ASM_clobbers(b);
3791 for (i = 0; i < n; ++i) {
3792 if (cla[i] != clb[i])
3796 } /* node_cmp_attr_ASM */
3799 * Set the default node attribute compare operation for an ir_op_ops.
3801 * @param code the opcode for the default operation
3802 * @param ops the operations initialized
3807 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
3811 ops->node_cmp_attr = node_cmp_attr_##a; \
3836 } /* firm_set_default_node_cmp_attr */
3839 * Compare function for two nodes in the hash table. Gets two
3840 * nodes as parameters. Returns 0 if the nodes are a cse.
3842 int identities_cmp(const void *elt, const void *key) {
3849 if (a == b) return 0;
3851 if ((get_irn_op(a) != get_irn_op(b)) ||
3852 (get_irn_mode(a) != get_irn_mode(b))) return 1;
3854 /* compare if a's in and b's in are of equal length */
3855 irn_arity_a = get_irn_intra_arity (a);
3856 if (irn_arity_a != get_irn_intra_arity(b))
3859 /* for block-local cse and op_pin_state_pinned nodes: */
3860 if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) {
3861 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
3865 /* compare a->in[0..ins] with b->in[0..ins] */
3866 for (i = 0; i < irn_arity_a; i++)
3867 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
3871 * here, we already now that the nodes are identical except their
3874 if (a->op->ops.node_cmp_attr)
3875 return a->op->ops.node_cmp_attr(a, b);
3878 } /* identities_cmp */
3881 * Calculate a hash value of a node.
3883 unsigned ir_node_hash(ir_node *node) {
3887 if (node->op == op_Const) {
3888 /* special value for const, as they only differ in their tarval. */
3889 h = HASH_PTR(node->attr.con.tv);
3890 h = 9*h + HASH_PTR(get_irn_mode(node));
3891 } else if (node->op == op_SymConst) {
3892 /* special value for const, as they only differ in their symbol. */
3893 h = HASH_PTR(node->attr.symc.sym.type_p);
3894 h = 9*h + HASH_PTR(get_irn_mode(node));
3897 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
3898 h = irn_arity = get_irn_intra_arity(node);
3900 /* consider all in nodes... except the block if not a control flow. */
3901 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
3902 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
3906 h = 9*h + HASH_PTR(get_irn_mode(node));
3908 h = 9*h + HASH_PTR(get_irn_op(node));
3912 } /* ir_node_hash */
3914 pset *new_identities(void) {
3915 return new_pset(identities_cmp, N_IR_NODES);
3916 } /* new_identities */
3918 void del_identities(pset *value_table) {
3919 del_pset(value_table);
3920 } /* del_identities */
3923 * Return the canonical node computing the same value as n.
3925 * @param value_table The value table
3926 * @param n The node to lookup
3928 * Looks up the node in a hash table.
3930 * For Const nodes this is performed in the constructor, too. Const
3931 * nodes are extremely time critical because of their frequent use in
3932 * constant string arrays.
3934 static INLINE ir_node *identify(pset *value_table, ir_node *n) {
3937 if (!value_table) return n;
3939 if (get_opt_reassociation()) {
3940 if (is_op_commutative(get_irn_op(n))) {
3941 ir_node *l = get_binop_left(n);
3942 ir_node *r = get_binop_right(n);
3944 /* for commutative operators perform a OP b == b OP a */
3945 if (get_irn_idx(l) > get_irn_idx(r)) {
3946 set_binop_left(n, r);
3947 set_binop_right(n, l);
3952 o = pset_find(value_table, n, ir_node_hash(n));
3961 * During construction we set the op_pin_state_pinned flag in the graph right when the
3962 * optimization is performed. The flag turning on procedure global cse could
3963 * be changed between two allocations. This way we are safe.
3965 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
3968 n = identify(value_table, n);
3969 if (get_irn_n(old, -1) != get_irn_n(n, -1))
3970 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3972 } /* identify_cons */
3975 * Return the canonical node computing the same value as n.
3976 * Looks up the node in a hash table, enters it in the table
3977 * if it isn't there yet.
3979 ir_node *identify_remember(pset *value_table, ir_node *n) {
3982 if (!value_table) return n;
3984 if (get_opt_reassociation()) {
3985 if (is_op_commutative(get_irn_op(n))) {
3986 ir_node *l = get_binop_left(n);
3987 ir_node *r = get_binop_right(n);
3988 int l_idx = get_irn_idx(l);
3989 int r_idx = get_irn_idx(r);
3991 /* For commutative operators perform a OP b == b OP a but keep
3992 constants on the RIGHT side. This helps greatly in some optimizations.
3993 Moreover we use the idx number to make the form deterministic. */
3994 if (is_irn_constlike(l))
3996 if (is_irn_constlike(r))
3998 if (l_idx < r_idx) {
3999 set_binop_left(n, r);
4000 set_binop_right(n, l);
4005 /* lookup or insert in hash table with given hash key. */
4006 o = pset_insert(value_table, n, ir_node_hash(n));
4013 } /* identify_remember */
4015 /* Add a node to the identities value table. */
4016 void add_identities(pset *value_table, ir_node *node) {
4017 if (get_opt_cse() && is_no_Block(node))
4018 identify_remember(value_table, node);
4019 } /* add_identities */
4021 /* Visit each node in the value table of a graph. */
4022 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
4024 ir_graph *rem = current_ir_graph;
4026 current_ir_graph = irg;
4027 foreach_pset(irg->value_table, node)
4029 current_ir_graph = rem;
4030 } /* visit_all_identities */
4033 * Garbage in, garbage out. If a node has a dead input, i.e., the
4034 * Bad node is input to the node, return the Bad node.
4036 static INLINE ir_node *gigo(ir_node *node) {
4038 ir_op *op = get_irn_op(node);
4040 /* remove garbage blocks by looking at control flow that leaves the block
4041 and replacing the control flow by Bad. */
4042 if (get_irn_mode(node) == mode_X) {
4043 ir_node *block = get_nodes_block(skip_Proj(node));
4045 /* Don't optimize nodes in immature blocks. */
4046 if (!get_Block_matured(block)) return node;
4047 /* Don't optimize End, may have Bads. */
4048 if (op == op_End) return node;
4050 if (is_Block(block)) {
4051 irn_arity = get_irn_arity(block);
4052 for (i = 0; i < irn_arity; i++) {
4053 if (!is_Bad(get_irn_n(block, i)))
4056 if (i == irn_arity) {
4057 ir_graph *irg = get_irn_irg(block);
4058 /* the start block is never dead */
4059 if (block != get_irg_start_block(irg)
4060 && block != get_irg_end_block(irg))
4066 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
4067 blocks predecessors is dead. */
4068 if (op != op_Block && op != op_Phi && op != op_Tuple) {
4069 irn_arity = get_irn_arity(node);
4072 * Beware: we can only read the block of a non-floating node.
4074 if (is_irn_pinned_in_irg(node) &&
4075 is_Block_dead(get_nodes_block(node)))
4078 for (i = 0; i < irn_arity; i++) {
4079 ir_node *pred = get_irn_n(node, i);
4084 /* Propagating Unknowns here seems to be a bad idea, because
4085 sometimes we need a node as a input and did not want that
4087 However, it might be useful to move this into a later phase
4088 (if you think that optimizing such code is useful). */
4089 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
4090 return new_Unknown(get_irn_mode(node));
4095 /* With this code we violate the agreement that local_optimize
4096 only leaves Bads in Block, Phi and Tuple nodes. */
4097 /* If Block has only Bads as predecessors it's garbage. */
4098 /* If Phi has only Bads as predecessors it's garbage. */
4099 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
4100 irn_arity = get_irn_arity(node);
4101 for (i = 0; i < irn_arity; i++) {
4102 if (!is_Bad(get_irn_n(node, i))) break;
4104 if (i == irn_arity) node = new_Bad();
4111 * These optimizations deallocate nodes from the obstack.
4112 * It can only be called if it is guaranteed that no other nodes
4113 * reference this one, i.e., right after construction of a node.
4115 * @param n The node to optimize
4117 * current_ir_graph must be set to the graph of the node!
4119 ir_node *optimize_node(ir_node *n) {
4122 ir_opcode iro = get_irn_opcode(n);
4124 /* Always optimize Phi nodes: part of the construction. */
4125 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
4127 /* constant expression evaluation / constant folding */
4128 if (get_opt_constant_folding()) {
4129 /* neither constants nor Tuple values can be evaluated */
4130 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
4131 /* try to evaluate */
4132 tv = computed_value(n);
4133 if (tv != tarval_bad) {
4135 ir_type *old_tp = get_irn_type(n);
4136 int i, arity = get_irn_arity(n);
4140 * Try to recover the type of the new expression.
4142 for (i = 0; i < arity && !old_tp; ++i)
4143 old_tp = get_irn_type(get_irn_n(n, i));
4146 * we MUST copy the node here temporary, because it's still needed
4147 * for DBG_OPT_CSTEVAL
4149 node_size = offsetof(ir_node, attr) + n->op->attr_size;
4150 oldn = alloca(node_size);
4152 memcpy(oldn, n, node_size);
4153 CLONE_ARR_A(ir_node *, oldn->in, n->in);
4155 /* ARG, copy the in array, we need it for statistics */
4156 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
4158 /* note the inplace edges module */
4159 edges_node_deleted(n, current_ir_graph);
4161 /* evaluation was successful -- replace the node. */
4162 irg_kill_node(current_ir_graph, n);
4163 nw = new_Const(get_tarval_mode (tv), tv);
4165 if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv))
4166 set_Const_type(nw, old_tp);
4167 DBG_OPT_CSTEVAL(oldn, nw);
4173 /* remove unnecessary nodes */
4174 if (get_opt_constant_folding() ||
4175 (iro == iro_Phi) || /* always optimize these nodes. */
4177 (iro == iro_Proj) ||
4178 (iro == iro_Block) ) /* Flags tested local. */
4179 n = equivalent_node(n);
4181 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
4183 /* Common Subexpression Elimination.
4185 * Checks whether n is already available.
4186 * The block input is used to distinguish different subexpressions. Right
4187 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
4188 * subexpressions within a block.
4191 n = identify_cons(current_ir_graph->value_table, n);
4194 edges_node_deleted(oldn, current_ir_graph);
4196 /* We found an existing, better node, so we can deallocate the old node. */
4197 irg_kill_node(current_ir_graph, oldn);
4201 /* Some more constant expression evaluation that does not allow to
4203 iro = get_irn_opcode(n);
4204 if (get_opt_constant_folding() ||
4205 (iro == iro_Cond) ||
4206 (iro == iro_Proj)) /* Flags tested local. */
4207 n = transform_node(n);
4209 /* Remove nodes with dead (Bad) input.
4210 Run always for transformation induced Bads. */
4213 /* Now we have a legal, useful node. Enter it in hash table for CSE */
4214 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
4215 n = identify_remember(current_ir_graph->value_table, n);
4219 } /* optimize_node */
4223 * These optimizations never deallocate nodes (in place). This can cause dead
4224 * nodes lying on the obstack. Remove these by a dead node elimination,
4225 * i.e., a copying garbage collection.
4227 ir_node *optimize_in_place_2(ir_node *n) {
4230 ir_opcode iro = get_irn_opcode(n);
4232 if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n;
4234 /* constant expression evaluation / constant folding */
4235 if (get_opt_constant_folding()) {
4236 /* neither constants nor Tuple values can be evaluated */
4237 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
4238 /* try to evaluate */
4239 tv = computed_value(n);
4240 if (tv != tarval_bad) {
4241 /* evaluation was successful -- replace the node. */
4242 ir_type *old_tp = get_irn_type(n);
4243 int i, arity = get_irn_arity(n);
4246 * Try to recover the type of the new expression.
4248 for (i = 0; i < arity && !old_tp; ++i)
4249 old_tp = get_irn_type(get_irn_n(n, i));
4251 n = new_Const(get_tarval_mode(tv), tv);
4253 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
4254 set_Const_type(n, old_tp);
4256 DBG_OPT_CSTEVAL(oldn, n);
4262 /* remove unnecessary nodes */
4263 if (get_opt_constant_folding() ||
4264 (iro == iro_Phi) || /* always optimize these nodes. */
4265 (iro == iro_Id) || /* ... */
4266 (iro == iro_Proj) || /* ... */
4267 (iro == iro_Block) ) /* Flags tested local. */
4268 n = equivalent_node(n);
4270 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
4272 /** common subexpression elimination **/
4273 /* Checks whether n is already available. */
4274 /* The block input is used to distinguish different subexpressions. Right
4275 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
4276 subexpressions within a block. */
4277 if (get_opt_cse()) {
4278 n = identify(current_ir_graph->value_table, n);
4281 /* Some more constant expression evaluation. */
4282 iro = get_irn_opcode(n);
4283 if (get_opt_constant_folding() ||
4284 (iro == iro_Cond) ||
4285 (iro == iro_Proj)) /* Flags tested local. */
4286 n = transform_node(n);
4288 /* Remove nodes with dead (Bad) input.
4289 Run always for transformation induced Bads. */
4292 /* Now we can verify the node, as it has no dead inputs any more. */
4295 /* Now we have a legal, useful node. Enter it in hash table for cse.
4296 Blocks should be unique anyways. (Except the successor of start:
4297 is cse with the start block!) */
4298 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
4299 n = identify_remember(current_ir_graph->value_table, n);
4302 } /* optimize_in_place_2 */
4305 * Wrapper for external use, set proper status bits after optimization.
4307 ir_node *optimize_in_place(ir_node *n) {
4308 /* Handle graph state */
4309 assert(get_irg_phase_state(current_ir_graph) != phase_building);
4311 if (get_opt_global_cse())
4312 set_irg_pinned(current_ir_graph, op_pin_state_floats);
4313 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
4314 set_irg_outs_inconsistent(current_ir_graph);
4316 /* FIXME: Maybe we could also test whether optimizing the node can
4317 change the control graph. */
4318 set_irg_doms_inconsistent(current_ir_graph);
4319 return optimize_in_place_2(n);
4320 } /* optimize_in_place */
4323 * Return the block for all default nodes.
4325 static ir_node *get_block_default(const ir_node *self) {
4326 return get_irn_n(self, -1);
4330 * Sets the block for all default nodes.
4332 static void set_block_default(ir_node *self, ir_node *blk) {
4333 set_irn_n(self, -1, blk);
4337 * It's not allowed to get the block of a block. Anyway, returns
4338 * the macroblock header in release mode.
4340 static ir_node *get_block_Block(const ir_node *self) {
4341 assert(!"get_nodes_block() called for a block");
4342 return get_irn_n(self, -1);
4346 * It's not allowed to set the block of a block. In release mode sets
4347 * the macroblock header.
4349 static void set_block_Block(ir_node *self, ir_node *blk) {
4350 assert(!"set_nodes_block() called for a block");
4351 set_irn_n(self, -1, blk);
4355 * The anchor is always placed in the endblock or a graph.
4357 static ir_node *get_block_Anchor(const ir_node *self) {
4358 return get_irn_n(self, anchor_end_block);
4362 * It's forbidden to set the anchor block.
4364 static void set_block_Anchor(ir_node *self, ir_node *blk) {
4367 assert(!"set_nodes_block() called for the Anchor");
4371 * Proj nodes are always in the block of it's predecessor.
4373 static ir_node *get_block_Proj(const ir_node *self) {
4374 ir_node *pred = get_Proj_pred(self);
4375 return get_nodes_block(pred);
4379 * Proj nodes silently ignore the block set request.
4381 static void set_block_Proj(ir_node *self, ir_node *blk) {
4387 * Set the default get_block operation.
4389 static ir_op_ops *firm_set_default_get_block(ir_opcode code, ir_op_ops *ops) {
4392 ops->get_block = get_block_##a; \
4393 ops->set_block = set_block_##a; \
4399 #ifndef CAN_PLACE_PROJS
4403 /* not allowed to be NULL */
4404 if (! ops->get_block)
4405 ops->get_block = get_block_default;
4406 if (! ops->set_block)
4407 ops->set_block = set_block_default;
4412 } /* firm_set_default_get_block */
4415 * Sets the default operation for an ir_ops.
4417 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
4418 ops = firm_set_default_get_block(code, ops);
4419 ops = firm_set_default_computed_value(code, ops);
4420 ops = firm_set_default_equivalent_node(code, ops);
4421 ops = firm_set_default_transform_node(code, ops);
4422 ops = firm_set_default_node_cmp_attr(code, ops);
4423 ops = firm_set_default_get_type(code, ops);
4424 ops = firm_set_default_get_type_attr(code, ops);
4425 ops = firm_set_default_get_entity_attr(code, ops);
4428 } /* firm_set_default_operations */