2 * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief iropt --- optimizations intertwined with IR construction.
23 * @author Christian Schaefer, Goetz Lindenmaier, Michael Beck
33 #include "irgraph_t.h"
34 #include "iredges_t.h"
41 #include "dbginfo_t.h"
42 #include "iropt_dbg.h"
48 #include "opt_confirms.h"
49 #include "opt_polymorphy.h"
53 /* Make types visible to allow most efficient access */
57 * Return the value of a Constant.
59 static tarval *computed_value_Const(ir_node *n) {
60 return get_Const_tarval(n);
61 } /* computed_value_Const */
64 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
66 static tarval *computed_value_SymConst(ir_node *n) {
70 switch (get_SymConst_kind(n)) {
71 case symconst_type_size:
72 type = get_SymConst_type(n);
73 if (get_type_state(type) == layout_fixed)
74 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
76 case symconst_type_align:
77 type = get_SymConst_type(n);
78 if (get_type_state(type) == layout_fixed)
79 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
81 case symconst_ofs_ent:
82 ent = get_SymConst_entity(n);
83 type = get_entity_owner(ent);
84 if (get_type_state(type) == layout_fixed)
85 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
91 } /* computed_value_SymConst */
94 * Return the value of an Add.
96 static tarval *computed_value_Add(ir_node *n) {
97 ir_node *a = get_Add_left(n);
98 ir_node *b = get_Add_right(n);
100 tarval *ta = value_of(a);
101 tarval *tb = value_of(b);
103 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b)))
104 return tarval_add(ta, tb);
107 } /* computed_value_Add */
110 * Return the value of a Sub.
111 * Special case: a - a
113 static tarval *computed_value_Sub(ir_node *n) {
114 ir_node *a = get_Sub_left(n);
115 ir_node *b = get_Sub_right(n);
120 if (a == b && !is_Bad(a))
121 return get_mode_null(get_irn_mode(n));
126 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b)))
127 return tarval_sub(ta, tb);
130 } /* computed_value_Sub */
133 * Return the value of a Carry.
134 * Special : a op 0, 0 op b
136 static tarval *computed_value_Carry(ir_node *n) {
137 ir_node *a = get_binop_left(n);
138 ir_node *b = get_binop_right(n);
139 ir_mode *m = get_irn_mode(n);
141 tarval *ta = value_of(a);
142 tarval *tb = value_of(b);
144 if ((ta != tarval_bad) && (tb != tarval_bad)) {
146 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
148 if ( (classify_tarval(ta) == TV_CLASSIFY_NULL)
149 || (classify_tarval(tb) == TV_CLASSIFY_NULL))
150 return get_mode_null(m);
153 } /* computed_value_Carry */
156 * Return the value of a Borrow.
159 static tarval *computed_value_Borrow(ir_node *n) {
160 ir_node *a = get_binop_left(n);
161 ir_node *b = get_binop_right(n);
162 ir_mode *m = get_irn_mode(n);
164 tarval *ta = value_of(a);
165 tarval *tb = value_of(b);
167 if ((ta != tarval_bad) && (tb != tarval_bad)) {
168 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
169 } else if (classify_tarval(ta) == TV_CLASSIFY_NULL) {
170 return get_mode_null(m);
173 } /* computed_value_Borrow */
176 * Return the value of an unary Minus.
178 static tarval *computed_value_Minus(ir_node *n) {
179 ir_node *a = get_Minus_op(n);
180 tarval *ta = value_of(a);
182 if ((ta != tarval_bad) && mode_is_signed(get_irn_mode(a)))
183 return tarval_neg(ta);
186 } /* computed_value_Minus */
189 * Return the value of a Mul.
191 static tarval *computed_value_Mul(ir_node *n) {
192 ir_node *a = get_Mul_left(n);
193 ir_node *b = get_Mul_right(n);
195 tarval *ta = value_of(a);
196 tarval *tb = value_of(b);
198 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
199 return tarval_mul(ta, tb);
201 /* a*0 = 0 or 0*b = 0:
202 calls computed_value recursive and returns the 0 with proper
204 if ((ta != tarval_bad) && (ta == get_mode_null(get_tarval_mode(ta))))
206 if ((tb != tarval_bad) && (tb == get_mode_null(get_tarval_mode(tb))))
210 } /* computed_value_Mul */
213 * Return the value of a floating point Quot.
215 static tarval *computed_value_Quot(ir_node *n) {
216 ir_node *a = get_Quot_left(n);
217 ir_node *b = get_Quot_right(n);
219 tarval *ta = value_of(a);
220 tarval *tb = value_of(b);
222 /* This was missing in original implementation. Why? */
223 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
224 if (tb != get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */
225 return tarval_quo(ta, tb);
228 } /* computed_value_Quot */
231 * Calculate the value of an integer Div of two nodes.
232 * Special case: 0 / b
234 static tarval *do_computed_value_Div(ir_node *a, ir_node *b) {
235 tarval *ta = value_of(a);
236 tarval *tb = value_of(b);
238 /* Compute c1 / c2 or 0 / a, a != 0 */
239 if (ta != tarval_bad) {
240 if ((tb != tarval_bad) && (tb != get_mode_null(get_irn_mode(b)))) /* div by zero: return tarval_bad */
241 return tarval_div(ta, tb);
242 else if (ta == get_mode_null(get_tarval_mode(ta))) /* 0 / b == 0 */
246 } /* do_computed_value_Div */
249 * Return the value of an integer Div.
251 static tarval *computed_value_Div(ir_node *n) {
252 return do_computed_value_Div(get_Div_left(n), get_Div_right(n));
253 } /* computed_value_Div */
256 * Calculate the value of an integer Mod of two nodes.
257 * Special case: a % 1
259 static tarval *do_computed_value_Mod(ir_node *a, ir_node *b) {
260 tarval *ta = value_of(a);
261 tarval *tb = value_of(b);
263 /* Compute c1 % c2 or a % 1 */
264 if (tb != tarval_bad) {
265 if ((ta != tarval_bad) && (tb != get_mode_null(get_tarval_mode(tb)))) /* div by zero: return tarval_bad */
266 return tarval_mod(ta, tb);
267 else if (tb == get_mode_one(get_tarval_mode(tb))) /* x mod 1 == 0 */
268 return get_mode_null(get_irn_mode(a));
271 } /* do_computed_value_Mod */
274 * Return the value of an integer Mod.
276 static tarval *computed_value_Mod(ir_node *n) {
277 return do_computed_value_Mod(get_Mod_left(n), get_Mod_right(n));
278 } /* computed_value_Mod */
281 * Return the value of an Abs.
283 static tarval *computed_value_Abs(ir_node *n) {
284 ir_node *a = get_Abs_op(n);
285 tarval *ta = value_of(a);
287 if (ta != tarval_bad)
288 return tarval_abs(ta);
291 } /* computed_value_Abs */
294 * Return the value of an And.
295 * Special case: a & 0, 0 & b
297 static tarval *computed_value_And(ir_node *n) {
298 ir_node *a = get_And_left(n);
299 ir_node *b = get_And_right(n);
301 tarval *ta = value_of(a);
302 tarval *tb = value_of(b);
304 if ((ta != tarval_bad) && (tb != tarval_bad)) {
305 return tarval_and (ta, tb);
309 if ( (classify_tarval ((v = ta)) == TV_CLASSIFY_NULL)
310 || (classify_tarval ((v = tb)) == TV_CLASSIFY_NULL)) {
315 } /* computed_value_And */
318 * Return the value of an Or.
319 * Special case: a | 1...1, 1...1 | b
321 static tarval *computed_value_Or(ir_node *n) {
322 ir_node *a = get_Or_left(n);
323 ir_node *b = get_Or_right(n);
325 tarval *ta = value_of(a);
326 tarval *tb = value_of(b);
328 if ((ta != tarval_bad) && (tb != tarval_bad)) {
329 return tarval_or (ta, tb);
332 if ( (classify_tarval ((v = ta)) == TV_CLASSIFY_ALL_ONE)
333 || (classify_tarval ((v = tb)) == TV_CLASSIFY_ALL_ONE)) {
338 } /* computed_value_Or */
341 * Return the value of an Eor.
343 static tarval *computed_value_Eor(ir_node *n) {
344 ir_node *a = get_Eor_left(n);
345 ir_node *b = get_Eor_right(n);
350 return get_mode_null(get_irn_mode(n));
355 if ((ta != tarval_bad) && (tb != tarval_bad)) {
356 return tarval_eor (ta, tb);
359 } /* computed_value_Eor */
362 * Return the value of a Not.
364 static tarval *computed_value_Not(ir_node *n) {
365 ir_node *a = get_Not_op(n);
366 tarval *ta = value_of(a);
368 if (ta != tarval_bad)
369 return tarval_not(ta);
372 } /* computed_value_Not */
375 * Return the value of a Shl.
377 static tarval *computed_value_Shl(ir_node *n) {
378 ir_node *a = get_Shl_left(n);
379 ir_node *b = get_Shl_right(n);
381 tarval *ta = value_of(a);
382 tarval *tb = value_of(b);
384 if ((ta != tarval_bad) && (tb != tarval_bad)) {
385 return tarval_shl (ta, tb);
388 } /* computed_value_Shl */
391 * Return the value of a Shr.
393 static tarval *computed_value_Shr(ir_node *n) {
394 ir_node *a = get_Shr_left(n);
395 ir_node *b = get_Shr_right(n);
397 tarval *ta = value_of(a);
398 tarval *tb = value_of(b);
400 if ((ta != tarval_bad) && (tb != tarval_bad)) {
401 return tarval_shr (ta, tb);
404 } /* computed_value_Shr */
407 * Return the value of a Shrs.
409 static tarval *computed_value_Shrs(ir_node *n) {
410 ir_node *a = get_Shrs_left(n);
411 ir_node *b = get_Shrs_right(n);
413 tarval *ta = value_of(a);
414 tarval *tb = value_of(b);
416 if ((ta != tarval_bad) && (tb != tarval_bad)) {
417 return tarval_shrs (ta, tb);
420 } /* computed_value_Shrs */
423 * Return the value of a Rot.
425 static tarval *computed_value_Rot(ir_node *n) {
426 ir_node *a = get_Rot_left(n);
427 ir_node *b = get_Rot_right(n);
429 tarval *ta = value_of(a);
430 tarval *tb = value_of(b);
432 if ((ta != tarval_bad) && (tb != tarval_bad)) {
433 return tarval_rot (ta, tb);
436 } /* computed_value_Rot */
439 * Return the value of a Conv.
441 static tarval *computed_value_Conv(ir_node *n) {
442 ir_node *a = get_Conv_op(n);
443 tarval *ta = value_of(a);
445 if (ta != tarval_bad)
446 return tarval_convert_to(ta, get_irn_mode(n));
449 } /* computed_value_Conv */
452 * Return the value of a Proj(Cmp).
454 * This performs a first step of unreachable code elimination.
455 * Proj can not be computed, but folding a Cmp above the Proj here is
456 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
458 * There are several case where we can evaluate a Cmp node, see later.
460 static tarval *computed_value_Proj_Cmp(ir_node *n) {
461 ir_node *a = get_Proj_pred(n);
462 ir_node *aa = get_Cmp_left(a);
463 ir_node *ab = get_Cmp_right(a);
464 long proj_nr = get_Proj_proj(n);
467 * BEWARE: a == a is NOT always True for floating Point values, as
468 * NaN != NaN is defined, so we must check this here.
471 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
474 /* This is a trick with the bits used for encoding the Cmp
475 Proj numbers, the following statement is not the same:
476 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
477 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
480 tarval *taa = value_of(aa);
481 tarval *tab = value_of(ab);
482 ir_mode *mode = get_irn_mode(aa);
485 * The predecessors of Cmp are target values. We can evaluate
488 if ((taa != tarval_bad) && (tab != tarval_bad)) {
489 /* strange checks... */
490 pn_Cmp flags = tarval_cmp(taa, tab);
491 if (flags != pn_Cmp_False) {
492 return new_tarval_from_long (proj_nr & flags, mode_b);
495 /* for integer values, we can check against MIN/MAX */
496 else if (mode_is_int(mode)) {
497 /* MIN <=/> x. This results in true/false. */
498 if (taa == get_mode_min(mode)) {
499 /* a compare with the MIN value */
500 if (proj_nr == pn_Cmp_Le)
501 return get_tarval_b_true();
502 else if (proj_nr == pn_Cmp_Gt)
503 return get_tarval_b_false();
505 /* x >=/< MIN. This results in true/false. */
507 if (tab == get_mode_min(mode)) {
508 /* a compare with the MIN value */
509 if (proj_nr == pn_Cmp_Ge)
510 return get_tarval_b_true();
511 else if (proj_nr == pn_Cmp_Lt)
512 return get_tarval_b_false();
514 /* MAX >=/< x. This results in true/false. */
515 else if (taa == get_mode_max(mode)) {
516 if (proj_nr == pn_Cmp_Ge)
517 return get_tarval_b_true();
518 else if (proj_nr == pn_Cmp_Lt)
519 return get_tarval_b_false();
521 /* x <=/> MAX. This results in true/false. */
522 else if (tab == get_mode_max(mode)) {
523 if (proj_nr == pn_Cmp_Le)
524 return get_tarval_b_true();
525 else if (proj_nr == pn_Cmp_Gt)
526 return get_tarval_b_false();
530 * The predecessors are Allocs or (void*)(0) constants. Allocs never
531 * return NULL, they raise an exception. Therefore we can predict
535 ir_node *aaa = skip_Id(skip_Proj(aa));
536 ir_node *aba = skip_Id(skip_Proj(ab));
538 if ( ( (/* aa is ProjP and aaa is Alloc */
539 (get_irn_op(aa) == op_Proj)
540 && (mode_is_reference(get_irn_mode(aa)))
541 && (get_irn_op(aaa) == op_Alloc))
542 && ( (/* ab is NULL */
543 (get_irn_op(ab) == op_Const)
544 && (mode_is_reference(get_irn_mode(ab)))
545 && (get_Const_tarval(ab) == get_mode_null(get_irn_mode(ab))))
546 || (/* ab is other Alloc */
547 (get_irn_op(ab) == op_Proj)
548 && (mode_is_reference(get_irn_mode(ab)))
549 && (get_irn_op(aba) == op_Alloc)
551 || (/* aa is NULL and aba is Alloc */
552 (get_irn_op(aa) == op_Const)
553 && (mode_is_reference(get_irn_mode(aa)))
554 && (get_Const_tarval(aa) == get_mode_null(get_irn_mode(aa)))
555 && (get_irn_op(ab) == op_Proj)
556 && (mode_is_reference(get_irn_mode(ab)))
557 && (get_irn_op(aba) == op_Alloc)))
559 return new_tarval_from_long(proj_nr & pn_Cmp_Ne, mode_b);
562 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
563 } /* computed_value_Proj_Cmp */
566 * Return the value of a Proj, handle Proj(Cmp), Proj(Div), Proj(Mod),
567 * Proj(DivMod) and Proj(Quot).
569 static tarval *computed_value_Proj(ir_node *n) {
570 ir_node *a = get_Proj_pred(n);
573 switch (get_irn_opcode(a)) {
575 return computed_value_Proj_Cmp(n);
578 /* compute either the Div or the Mod part */
579 proj_nr = get_Proj_proj(n);
580 if (proj_nr == pn_DivMod_res_div)
581 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
582 else if (proj_nr == pn_DivMod_res_mod)
583 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
587 if (get_Proj_proj(n) == pn_Div_res)
588 return computed_value(a);
592 if (get_Proj_proj(n) == pn_Mod_res)
593 return computed_value(a);
597 if (get_Proj_proj(n) == pn_Quot_res)
598 return computed_value(a);
605 } /* computed_value_Proj */
608 * Calculate the value of a Mux: can be evaluated, if the
609 * sel and the right input are known.
611 static tarval *computed_value_Mux(ir_node *n) {
612 ir_node *sel = get_Mux_sel(n);
613 tarval *ts = value_of(sel);
615 if (ts == get_tarval_b_true()) {
616 ir_node *v = get_Mux_true(n);
619 else if (ts == get_tarval_b_false()) {
620 ir_node *v = get_Mux_false(n);
624 } /* computed_value_Mux */
627 * Calculate the value of a Psi: can be evaluated, if a condition is true
628 * and all previous conditions are false. If all conditions are false
629 * we evaluate to the default one.
631 static tarval *computed_value_Psi(ir_node *n) {
633 return computed_value_Mux(n);
635 } /* computed_value_Psi */
638 * Calculate the value of a Confirm: can be evaluated,
639 * if it has the form Confirm(x, '=', Const).
641 static tarval *computed_value_Confirm(ir_node *n) {
642 return get_Confirm_cmp(n) == pn_Cmp_Eq ?
643 value_of(get_Confirm_bound(n)) : tarval_bad;
644 } /* computed_value_Confirm */
647 * If the parameter n can be computed, return its value, else tarval_bad.
648 * Performs constant folding.
650 * @param n The node this should be evaluated
652 tarval *computed_value(ir_node *n) {
653 if (n->op->ops.computed_value)
654 return n->op->ops.computed_value(n);
656 } /* computed_value */
659 * Set the default computed_value evaluator in an ir_op_ops.
661 * @param code the opcode for the default operation
662 * @param ops the operations initialized
667 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
671 ops->computed_value = computed_value_##a; \
706 } /* firm_set_default_computed_value */
709 * Returns a equivalent block for another block.
710 * If the block has only one predecessor, this is
711 * the equivalent one. If the only predecessor of a block is
712 * the block itself, this is a dead block.
714 * If both predecessors of a block are the branches of a binary
715 * Cond, the equivalent block is Cond's block.
717 * If all predecessors of a block are bad or lies in a dead
718 * block, the current block is dead as well.
720 * Note, that blocks are NEVER turned into Bad's, instead
721 * the dead_block flag is set. So, never test for is_Bad(block),
722 * always use is_dead_Block(block).
724 static ir_node *equivalent_node_Block(ir_node *n)
727 int n_preds = get_Block_n_cfgpreds(n);
729 /* The Block constructor does not call optimize, but mature_immBlock
730 calls the optimization. */
731 assert(get_Block_matured(n));
733 /* Straightening: a single entry Block following a single exit Block
734 can be merged, if it is not the Start block. */
735 /* !!! Beware, all Phi-nodes of n must have been optimized away.
736 This should be true, as the block is matured before optimize is called.
737 But what about Phi-cycles with the Phi0/Id that could not be resolved?
738 Remaining Phi nodes are just Ids. */
739 if ((n_preds == 1) && (get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp)) {
740 ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0));
741 if (predblock == oldn) {
742 /* Jmp jumps into the block it is in -- deal self cycle. */
743 n = set_Block_dead(n);
744 DBG_OPT_DEAD_BLOCK(oldn, n);
745 } else if (get_opt_control_flow_straightening()) {
747 DBG_OPT_STG(oldn, n);
749 } else if ((n_preds == 1) &&
750 (get_irn_op(skip_Proj(get_Block_cfgpred(n, 0))) == op_Cond)) {
751 ir_node *predblock = get_Block_cfgpred_block(n, 0);
752 if (predblock == oldn) {
753 /* Jmp jumps into the block it is in -- deal self cycle. */
754 n = set_Block_dead(n);
755 DBG_OPT_DEAD_BLOCK(oldn, n);
757 } else if ((n_preds == 2) &&
758 (get_opt_control_flow_weak_simplification())) {
759 /* Test whether Cond jumps twice to this block
760 * The more general case which more than 2 predecessors is handles
761 * in optimize_cf(), we handle only this special case for speed here.
763 ir_node *a = get_Block_cfgpred(n, 0);
764 ir_node *b = get_Block_cfgpred(n, 1);
766 if ((get_irn_op(a) == op_Proj) &&
767 (get_irn_op(b) == op_Proj) &&
768 (get_Proj_pred(a) == get_Proj_pred(b)) &&
769 (get_irn_op(get_Proj_pred(a)) == op_Cond) &&
770 (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) {
771 /* Also a single entry Block following a single exit Block. Phis have
772 twice the same operand and will be optimized away. */
773 n = get_nodes_block(get_Proj_pred(a));
774 DBG_OPT_IFSIM1(oldn, a, b, n);
776 } else if (get_opt_unreachable_code() &&
777 (n != get_irg_start_block(current_ir_graph)) &&
778 (n != get_irg_end_block(current_ir_graph)) ) {
781 /* If all inputs are dead, this block is dead too, except if it is
782 the start or end block. This is one step of unreachable code
784 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
785 ir_node *pred = get_Block_cfgpred(n, i);
788 if (is_Bad(pred)) continue;
789 pred_blk = get_nodes_block(skip_Proj(pred));
791 if (is_Block_dead(pred_blk)) continue;
794 /* really found a living input */
799 n = set_Block_dead(n);
800 DBG_OPT_DEAD_BLOCK(oldn, n);
805 } /* equivalent_node_Block */
808 * Returns a equivalent node for a Jmp, a Bad :-)
809 * Of course this only happens if the Block of the Jmp is dead.
811 static ir_node *equivalent_node_Jmp(ir_node *n) {
812 /* unreachable code elimination */
813 if (is_Block_dead(get_nodes_block(n)))
817 } /* equivalent_node_Jmp */
819 /** Raise is handled in the same way as Jmp. */
820 #define equivalent_node_Raise equivalent_node_Jmp
823 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
824 See transform_node_Proj_Cond(). */
827 * Optimize operations that are commutative and have neutral 0,
828 * so a op 0 = 0 op a = a.
830 static ir_node *equivalent_node_neutral_zero(ir_node *n)
834 ir_node *a = get_binop_left(n);
835 ir_node *b = get_binop_right(n);
840 /* After running compute_node there is only one constant predecessor.
841 Find this predecessors value and remember the other node: */
842 if ((tv = value_of(a)) != tarval_bad) {
844 } else if ((tv = value_of(b)) != tarval_bad) {
849 /* If this predecessors constant value is zero, the operation is
850 * unnecessary. Remove it.
852 * Beware: If n is a Add, the mode of on and n might be different
853 * which happens in this rare construction: NULL + 3.
854 * Then, a Conv would be needed which we cannot include here.
856 if (classify_tarval (tv) == TV_CLASSIFY_NULL) {
857 if (get_irn_mode(on) == get_irn_mode(n)) {
860 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
865 } /* equivalent_node_neutral_zero */
868 * Eor is commutative and has neutral 0.
870 #define equivalent_node_Eor equivalent_node_neutral_zero
873 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
875 * The second one looks strange, but this construct
876 * is used heavily in the LCC sources :-).
878 * Beware: The Mode of an Add may be different than the mode of its
879 * predecessors, so we could not return a predecessors in all cases.
881 static ir_node *equivalent_node_Add(ir_node *n) {
883 ir_node *left, *right;
884 ir_mode *mode = get_irn_mode(n);
886 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
887 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
890 n = equivalent_node_neutral_zero(n);
894 left = get_Add_left(n);
895 right = get_Add_right(n);
897 if (get_irn_op(left) == op_Sub) {
898 if (get_Sub_right(left) == right) {
901 n = get_Sub_left(left);
902 if (mode == get_irn_mode(n)) {
903 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
908 if (get_irn_op(right) == op_Sub) {
909 if (get_Sub_right(right) == left) {
912 n = get_Sub_left(right);
913 if (mode == get_irn_mode(n)) {
914 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
920 } /* equivalent_node_Add */
923 * optimize operations that are not commutative but have neutral 0 on left,
926 static ir_node *equivalent_node_left_zero(ir_node *n) {
929 ir_node *a = get_binop_left(n);
930 ir_node *b = get_binop_right(n);
932 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
935 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
938 } /* equivalent_node_left_zero */
940 #define equivalent_node_Shl equivalent_node_left_zero
941 #define equivalent_node_Shr equivalent_node_left_zero
942 #define equivalent_node_Shrs equivalent_node_left_zero
943 #define equivalent_node_Rot equivalent_node_left_zero
946 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
948 * The second one looks strange, but this construct
949 * is used heavily in the LCC sources :-).
951 * Beware: The Mode of a Sub may be different than the mode of its
952 * predecessors, so we could not return a predecessors in all cases.
954 static ir_node *equivalent_node_Sub(ir_node *n) {
957 ir_mode *mode = get_irn_mode(n);
959 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
960 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
963 b = get_Sub_right(n);
965 /* Beware: modes might be different */
966 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
967 ir_node *a = get_Sub_left(n);
968 if (mode == get_irn_mode(a)) {
971 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
975 } /* equivalent_node_Sub */
979 * Optimize an "idempotent unary op", ie op(op(n)) = n.
982 * -(-a) == a, but might overflow two times.
983 * We handle it anyway here but the better way would be a
984 * flag. This would be needed for Pascal for instance.
986 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
988 ir_node *pred = get_unop_op(n);
990 /* optimize symmetric unop */
991 if (get_irn_op(pred) == get_irn_op(n)) {
992 n = get_unop_op(pred);
993 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
996 } /* equivalent_node_idempotent_unop */
998 /** Optimize Not(Not(x)) == x. */
999 #define equivalent_node_Not equivalent_node_idempotent_unop
1001 /** -(-x) == x ??? Is this possible or can --x raise an
1002 out of bounds exception if min =! max? */
1003 #define equivalent_node_Minus equivalent_node_idempotent_unop
1006 * Optimize a * 1 = 1 * a = a.
1008 static ir_node *equivalent_node_Mul(ir_node *n) {
1010 ir_node *a = get_Mul_left(n);
1011 ir_node *b = get_Mul_right(n);
1013 /* Mul is commutative and has again an other neutral element. */
1014 if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) {
1016 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1017 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) {
1019 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1022 } /* equivalent_node_Mul */
1025 * Optimize a / 1 = a.
1027 static ir_node *equivalent_node_Div(ir_node *n) {
1028 ir_node *a = get_Div_left(n);
1029 ir_node *b = get_Div_right(n);
1031 /* Div is not commutative. */
1032 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1033 /* Turn Div into a tuple (mem, bad, a) */
1034 ir_node *mem = get_Div_mem(n);
1035 ir_node *blk = get_irn_n(n, -1);
1036 turn_into_tuple(n, pn_Div_max);
1037 set_Tuple_pred(n, pn_Div_M, mem);
1038 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
1039 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
1040 set_Tuple_pred(n, pn_Div_res, a);
1043 } /* equivalent_node_Div */
1046 * Optimize a / 1.0 = a.
1048 static ir_node *equivalent_node_Quot(ir_node *n) {
1049 ir_node *a = get_Quot_left(n);
1050 ir_node *b = get_Quot_right(n);
1052 /* Div is not commutative. */
1053 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* Quot(x, 1) == x */
1054 /* Turn Quot into a tuple (mem, jmp, bad, a) */
1055 ir_node *mem = get_Quot_mem(n);
1056 ir_node *blk = get_irn_n(n, -1);
1057 turn_into_tuple(n, pn_Quot_max);
1058 set_Tuple_pred(n, pn_Quot_M, mem);
1059 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
1060 set_Tuple_pred(n, pn_Quot_X_except, new_Bad()); /* no exception */
1061 set_Tuple_pred(n, pn_Quot_res, a);
1064 } /* equivalent_node_Quot */
1067 * Optimize a / 1 = a.
1069 static ir_node *equivalent_node_DivMod(ir_node *n) {
1070 ir_node *b = get_DivMod_right(n);
1072 /* Div is not commutative. */
1073 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1074 /* Turn DivMod into a tuple (mem, jmp, bad, a, 0) */
1075 ir_node *a = get_DivMod_left(n);
1076 ir_node *mem = get_Div_mem(n);
1077 ir_node *blk = get_irn_n(n, -1);
1078 ir_mode *mode = get_DivMod_resmode(n);
1080 turn_into_tuple(n, pn_DivMod_max);
1081 set_Tuple_pred(n, pn_DivMod_M, mem);
1082 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
1083 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1084 set_Tuple_pred(n, pn_DivMod_res_div, a);
1085 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1088 } /* equivalent_node_DivMod */
1091 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1093 static ir_node *equivalent_node_Or(ir_node *n) {
1096 ir_node *a = get_Or_left(n);
1097 ir_node *b = get_Or_right(n);
1100 n = a; /* Or has it's own neutral element */
1101 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1102 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) {
1104 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1105 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
1107 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1111 } /* equivalent_node_Or */
1114 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a.
1116 static ir_node *equivalent_node_And(ir_node *n) {
1119 ir_node *a = get_And_left(n);
1120 ir_node *b = get_And_right(n);
1123 n = a; /* And has it's own neutral element */
1124 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1125 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) {
1127 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1128 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) {
1130 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1133 } /* equivalent_node_And */
1136 * Try to remove useless Conv's:
1138 static ir_node *equivalent_node_Conv(ir_node *n) {
1140 ir_node *a = get_Conv_op(n);
1143 ir_mode *n_mode = get_irn_mode(n);
1144 ir_mode *a_mode = get_irn_mode(a);
1146 if (n_mode == a_mode) { /* No Conv necessary */
1147 if (get_Conv_strict(n)) {
1148 /* special case: the predecessor might be a also a Conv */
1150 if (! get_Conv_strict(a)) {
1151 /* first one is not strict, kick it */
1152 set_Conv_op(n, get_Conv_op(a));
1155 /* else both are strict conv, second is superflous */
1157 /* leave strict floating point Conv's */
1162 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1163 } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */
1167 n_mode = get_irn_mode(n);
1168 b_mode = get_irn_mode(b);
1170 if (n_mode == b_mode) {
1171 if (n_mode == mode_b) {
1172 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1173 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1174 } else if (mode_is_int(n_mode)) {
1175 if (smaller_mode(b_mode, a_mode)){
1176 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1177 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1183 } /* equivalent_node_Conv */
1186 * A Cast may be removed if the type of the previous node
1187 * is already the type of the Cast.
1189 static ir_node *equivalent_node_Cast(ir_node *n) {
1191 ir_node *pred = get_Cast_op(n);
1193 if (get_irn_type(pred) == get_Cast_type(n)) {
1195 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1198 } /* equivalent_node_Cast */
1201 * Several optimizations:
1202 * - no Phi in start block.
1203 * - remove Id operators that are inputs to Phi
1204 * - fold Phi-nodes, iff they have only one predecessor except
1207 static ir_node *equivalent_node_Phi(ir_node *n) {
1211 ir_node *block = NULL; /* to shutup gcc */
1212 ir_node *first_val = NULL; /* to shutup gcc */
1214 if (!get_opt_normalize()) return n;
1216 n_preds = get_Phi_n_preds(n);
1218 block = get_nodes_block(n);
1219 /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
1220 assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
1221 if ((is_Block_dead(block)) || /* Control dead */
1222 (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */
1223 return new_Bad(); /* in the Start Block. */
1225 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1227 /* If the Block has a Bad pred, we also have one. */
1228 for (i = 0; i < n_preds; ++i)
1229 if (is_Bad(get_Block_cfgpred(block, i)))
1230 set_Phi_pred(n, i, new_Bad());
1232 /* Find first non-self-referencing input */
1233 for (i = 0; i < n_preds; ++i) {
1234 first_val = get_Phi_pred(n, i);
1235 if ( (first_val != n) /* not self pointer */
1237 && (! is_Bad(first_val))
1239 ) { /* value not dead */
1240 break; /* then found first value. */
1245 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1249 /* search for rest of inputs, determine if any of these
1250 are non-self-referencing */
1251 while (++i < n_preds) {
1252 ir_node *scnd_val = get_Phi_pred(n, i);
1253 if ( (scnd_val != n)
1254 && (scnd_val != first_val)
1256 && (! is_Bad(scnd_val))
1264 /* Fold, if no multiple distinct non-self-referencing inputs */
1266 DBG_OPT_PHI(oldn, n);
1269 } /* equivalent_node_Phi */
1272 * Several optimizations:
1273 * - no Sync in start block.
1274 * - fold Sync-nodes, iff they have only one predecessor except
1277 static ir_node *equivalent_node_Sync(ir_node *n) {
1281 ir_node *first_val = NULL; /* to shutup gcc */
1283 if (!get_opt_normalize()) return n;
1285 n_preds = get_Sync_n_preds(n);
1287 /* Find first non-self-referencing input */
1288 for (i = 0; i < n_preds; ++i) {
1289 first_val = get_Sync_pred(n, i);
1290 if ((first_val != n) /* not self pointer */ &&
1291 (! is_Bad(first_val))
1292 ) { /* value not dead */
1293 break; /* then found first value. */
1298 /* A totally Bad or self-referencing Sync (we didn't break the above loop) */
1301 /* search the rest of inputs, determine if any of these
1302 are non-self-referencing */
1303 while (++i < n_preds) {
1304 ir_node *scnd_val = get_Sync_pred(n, i);
1305 if ((scnd_val != n) &&
1306 (scnd_val != first_val) &&
1307 (! is_Bad(scnd_val))
1313 /* Fold, if no multiple distinct non-self-referencing inputs */
1315 DBG_OPT_SYNC(oldn, n);
1318 } /* equivalent_node_Sync */
1321 * Optimize Proj(Tuple) and gigo() for ProjX in Bad block,
1322 * ProjX(Load) and ProjX(Store).
1324 static ir_node *equivalent_node_Proj(ir_node *proj) {
1325 ir_node *oldn = proj;
1326 ir_node *a = get_Proj_pred(proj);
1328 if (get_irn_op(a) == op_Tuple) {
1329 /* Remove the Tuple/Proj combination. */
1330 if ( get_Proj_proj(proj) <= get_Tuple_n_preds(a) ) {
1331 proj = get_Tuple_pred(a, get_Proj_proj(proj));
1332 DBG_OPT_TUPLE(oldn, a, proj);
1334 /* This should not happen! */
1335 assert(! "found a Proj with higher number than Tuple predecessors");
1338 } else if (get_irn_mode(proj) == mode_X) {
1339 if (is_Block_dead(get_nodes_block(skip_Proj(proj)))) {
1340 /* Remove dead control flow -- early gigo(). */
1342 } else if (get_opt_ldst_only_null_ptr_exceptions()) {
1343 ir_op *op = get_irn_op(a);
1345 if (op == op_Load) {
1346 /* get the Load address */
1347 ir_node *addr = get_Load_ptr(a);
1348 ir_node *blk = get_irn_n(a, -1);
1351 if (value_not_null(addr, &confirm)) {
1352 if (confirm == NULL) {
1353 /* this node may float if it did not depend on a Confirm */
1354 set_irn_pinned(a, op_pin_state_floats);
1356 if (get_Proj_proj(proj) == pn_Load_X_except) {
1357 DBG_OPT_EXC_REM(proj);
1360 return new_r_Jmp(current_ir_graph, blk);
1362 } else if (op == op_Store) {
1363 /* get the load/store address */
1364 ir_node *addr = get_Store_ptr(a);
1365 ir_node *blk = get_irn_n(a, -1);
1368 if (value_not_null(addr, &confirm)) {
1369 if (confirm == NULL) {
1370 /* this node may float if it did not depend on a Confirm */
1371 set_irn_pinned(a, op_pin_state_floats);
1373 if (get_Proj_proj(proj) == pn_Store_X_except) {
1374 DBG_OPT_EXC_REM(proj);
1377 return new_r_Jmp(current_ir_graph, blk);
1384 } /* equivalent_node_Proj */
1389 static ir_node *equivalent_node_Id(ir_node *n) {
1394 } while (get_irn_op(n) == op_Id);
1396 DBG_OPT_ID(oldn, n);
1398 } /* equivalent_node_Id */
1403 static ir_node *equivalent_node_Mux(ir_node *n)
1405 ir_node *oldn = n, *sel = get_Mux_sel(n);
1406 tarval *ts = value_of(sel);
1408 /* Mux(true, f, t) == t */
1409 if (ts == tarval_b_true) {
1410 n = get_Mux_true(n);
1411 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1413 /* Mux(false, f, t) == f */
1414 else if (ts == tarval_b_false) {
1415 n = get_Mux_false(n);
1416 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1418 /* Mux(v, x, x) == x */
1419 else if (get_Mux_false(n) == get_Mux_true(n)) {
1420 n = get_Mux_true(n);
1421 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1423 else if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(get_irn_mode(n))) {
1424 ir_node *cmp = get_Proj_pred(sel);
1425 long proj_nr = get_Proj_proj(sel);
1426 ir_node *b = get_Mux_false(n);
1427 ir_node *a = get_Mux_true(n);
1430 * Note: normalization puts the constant on the right site,
1431 * so we check only one case.
1433 * Note further that these optimization work even for floating point
1434 * with NaN's because -NaN == NaN.
1435 * However, if +0 and -0 is handled differently, we cannot use the first one.
1437 if (get_irn_op(cmp) == op_Cmp && get_Cmp_left(cmp) == a) {
1438 if (classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
1439 /* Mux(a CMP 0, X, a) */
1440 if (get_irn_op(b) == op_Minus && get_Minus_op(b) == a) {
1441 /* Mux(a CMP 0, -a, a) */
1442 if (proj_nr == pn_Cmp_Eq) {
1443 /* Mux(a == 0, -a, a) ==> -a */
1445 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1446 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1447 /* Mux(a != 0, -a, a) ==> a */
1449 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1451 } else if (classify_Const(b) == CNST_NULL) {
1452 /* Mux(a CMP 0, 0, a) */
1453 if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1454 /* Mux(a != 0, 0, a) ==> a */
1456 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1457 } else if (proj_nr == pn_Cmp_Eq) {
1458 /* Mux(a == 0, 0, a) ==> 0 */
1460 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1467 } /* equivalent_node_Mux */
1470 * Returns a equivalent node of a Psi: if a condition is true
1471 * and all previous conditions are false we know its value.
1472 * If all conditions are false its value is the default one.
1474 static ir_node *equivalent_node_Psi(ir_node *n) {
1476 return equivalent_node_Mux(n);
1478 } /* equivalent_node_Psi */
1481 * Optimize -a CMP -b into b CMP a.
1482 * This works only for for modes where unary Minus
1484 * Note that two-complement integers can Overflow
1485 * so it will NOT work.
1487 * For == and != can be handled in Proj(Cmp)
1489 static ir_node *equivalent_node_Cmp(ir_node *n) {
1490 ir_node *left = get_Cmp_left(n);
1491 ir_node *right = get_Cmp_right(n);
1493 if (get_irn_op(left) == op_Minus && get_irn_op(right) == op_Minus &&
1494 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1495 left = get_Minus_op(left);
1496 right = get_Minus_op(right);
1497 set_Cmp_left(n, right);
1498 set_Cmp_right(n, left);
1501 } /* equivalent_node_Cmp */
1504 * Remove Confirm nodes if setting is on.
1505 * Replace Confirms(x, '=', Constlike) by Constlike.
1507 static ir_node *equivalent_node_Confirm(ir_node *n) {
1508 ir_node *pred = get_Confirm_value(n);
1509 pn_Cmp pnc = get_Confirm_cmp(n);
1511 if (get_irn_op(pred) == op_Confirm && pnc == get_Confirm_cmp(pred)) {
1513 * rare case: two identical Confirms one after another,
1514 * replace the second one with the first.
1518 if (pnc == pn_Cmp_Eq) {
1519 ir_node *bound = get_Confirm_bound(n);
1522 * Optimize a rare case:
1523 * Confirm(x, '=', Constlike) ==> Constlike
1525 if (is_irn_constlike(bound)) {
1526 DBG_OPT_CONFIRM(n, bound);
1530 return get_opt_remove_confirm() ? get_Confirm_value(n) : n;
1534 * Optimize CopyB(mem, x, x) into a Nop.
1536 static ir_node *equivalent_node_CopyB(ir_node *n) {
1537 ir_node *a = get_CopyB_dst(n);
1538 ir_node *b = get_CopyB_src(n);
1541 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1542 ir_node *mem = get_CopyB_mem(n);
1543 ir_node *blk = get_nodes_block(n);
1544 turn_into_tuple(n, pn_CopyB_max);
1545 set_Tuple_pred(n, pn_CopyB_M, mem);
1546 set_Tuple_pred(n, pn_CopyB_X_regular, new_r_Jmp(current_ir_graph, blk));
1547 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1548 set_Tuple_pred(n, pn_CopyB_M_except, new_Bad());
1551 } /* equivalent_node_CopyB */
1554 * Optimize Bounds(idx, idx, upper) into idx.
1556 static ir_node *equivalent_node_Bound(ir_node *n) {
1557 ir_node *idx = get_Bound_index(n);
1558 ir_node *lower = get_Bound_lower(n);
1561 /* By definition lower < upper, so if idx == lower -->
1562 lower <= idx && idx < upper */
1564 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1567 ir_node *pred = skip_Proj(idx);
1569 if (get_irn_op(pred) == op_Bound) {
1571 * idx was Bounds_check previously, it is still valid if
1572 * lower <= pred_lower && pred_upper <= upper.
1574 ir_node *upper = get_Bound_upper(n);
1575 if (get_Bound_lower(pred) == lower &&
1576 get_Bound_upper(pred) == upper) {
1578 * One could expect that we simply return the previous
1579 * Bound here. However, this would be wrong, as we could
1580 * add an exception Proj to a new location then.
1581 * So, we must turn in into a tuple.
1588 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1589 ir_node *mem = get_Bound_mem(n);
1590 ir_node *blk = get_nodes_block(n);
1591 turn_into_tuple(n, pn_Bound_max);
1592 set_Tuple_pred(n, pn_Bound_M, mem);
1593 set_Tuple_pred(n, pn_Bound_X_regular, new_r_Jmp(current_ir_graph, blk)); /* no exception */
1594 set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */
1595 set_Tuple_pred(n, pn_Bound_res, idx);
1598 } /* equivalent_node_Bound */
1601 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1602 * perform no actual computation, as, e.g., the Id nodes. It does not create
1603 * new nodes. It is therefore safe to free n if the node returned is not n.
1604 * If a node returns a Tuple we can not just skip it. If the size of the
1605 * in array fits, we transform n into a tuple (e.g., Div).
1607 ir_node *equivalent_node(ir_node *n) {
1608 if (n->op->ops.equivalent_node)
1609 return n->op->ops.equivalent_node(n);
1611 } /* equivalent_node */
1614 * Sets the default equivalent node operation for an ir_op_ops.
1616 * @param code the opcode for the default operation
1617 * @param ops the operations initialized
1622 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1626 ops->equivalent_node = equivalent_node_##a; \
1666 } /* firm_set_default_equivalent_node */
1669 * Do node specific optimizations of nodes predecessors.
1671 static void optimize_preds(ir_node *n) {
1672 switch (get_irn_opcode(n)) {
1674 case iro_Cmp: { /* We don't want Cast as input to Cmp. */
1675 ir_node *a = get_Cmp_left(n), *b = get_Cmp_right(n);
1677 if (get_irn_op(a) == op_Cast) {
1681 if (get_irn_op(b) == op_Cast) {
1683 set_Cmp_right(n, b);
1690 } /* optimize_preds */
1693 * Returns non-zero if a node is a Phi node
1694 * with all predecessors constant.
1696 static int is_const_Phi(ir_node *n) {
1701 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1702 if (! is_Const(get_irn_n(n, i)))
1705 } /* is_const_Phi */
1708 * Apply an evaluator on a binop with a constant operators (and one Phi).
1710 * @param phi the Phi node
1711 * @param other the other operand
1712 * @param eval an evaluator function
1713 * @param left if non-zero, other is the left operand, else the right
1715 * @return a new Phi node if the conversion was successful, NULL else
1717 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), int left) {
1723 int i, n = get_irn_arity(phi);
1725 NEW_ARR_A(void *, res, n);
1727 for (i = 0; i < n; ++i) {
1728 pred = get_irn_n(phi, i);
1729 tv = get_Const_tarval(pred);
1730 tv = eval(other, tv);
1732 if (tv == tarval_bad) {
1733 /* folding failed, bad */
1739 for (i = 0; i < n; ++i) {
1740 pred = get_irn_n(phi, i);
1741 tv = get_Const_tarval(pred);
1742 tv = eval(tv, other);
1744 if (tv == tarval_bad) {
1745 /* folding failed, bad */
1751 mode = get_irn_mode(phi);
1752 irg = current_ir_graph;
1753 for (i = 0; i < n; ++i) {
1754 pred = get_irn_n(phi, i);
1755 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1756 mode, res[i], get_Const_type(pred));
1758 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1759 } /* apply_binop_on_phi */
1762 * Apply an evaluator on a unop with a constant operator (a Phi).
1764 * @param phi the Phi node
1765 * @param eval an evaluator function
1767 * @return a new Phi node if the conversion was successful, NULL else
1769 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
1775 int i, n = get_irn_arity(phi);
1777 NEW_ARR_A(void *, res, n);
1778 for (i = 0; i < n; ++i) {
1779 pred = get_irn_n(phi, i);
1780 tv = get_Const_tarval(pred);
1783 if (tv == tarval_bad) {
1784 /* folding failed, bad */
1789 mode = get_irn_mode(phi);
1790 irg = current_ir_graph;
1791 for (i = 0; i < n; ++i) {
1792 pred = get_irn_n(phi, i);
1793 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1794 mode, res[i], get_Const_type(pred));
1796 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1797 } /* apply_unop_on_phi */
1800 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1801 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1802 * If possible, remove the Conv's.
1804 static ir_node *transform_node_AddSub(ir_node *n) {
1805 ir_mode *mode = get_irn_mode(n);
1807 if (mode_is_reference(mode)) {
1808 ir_node *left = get_binop_left(n);
1809 ir_node *right = get_binop_right(n);
1810 int ref_bits = get_mode_size_bits(mode);
1812 if (get_irn_op(left) == op_Conv) {
1813 ir_mode *mode = get_irn_mode(left);
1814 int bits = get_mode_size_bits(mode);
1816 if (ref_bits == bits &&
1817 mode_is_int(mode) &&
1818 get_mode_arithmetic(mode) == irma_twos_complement) {
1819 ir_node *pre = get_Conv_op(left);
1820 ir_mode *pre_mode = get_irn_mode(pre);
1822 if (mode_is_int(pre_mode) &&
1823 get_mode_size_bits(pre_mode) == bits &&
1824 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1825 /* ok, this conv just changes to sign, moreover the calculation
1826 * is done with same number of bits as our address mode, so
1827 * we can ignore the conv as address calculation can be viewed
1828 * as either signed or unsigned
1830 set_binop_left(n, pre);
1835 if (get_irn_op(right) == op_Conv) {
1836 ir_mode *mode = get_irn_mode(right);
1837 int bits = get_mode_size_bits(mode);
1839 if (ref_bits == bits &&
1840 mode_is_int(mode) &&
1841 get_mode_arithmetic(mode) == irma_twos_complement) {
1842 ir_node *pre = get_Conv_op(right);
1843 ir_mode *pre_mode = get_irn_mode(pre);
1845 if (mode_is_int(pre_mode) &&
1846 get_mode_size_bits(pre_mode) == bits &&
1847 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1848 /* ok, this conv just changes to sign, moreover the calculation
1849 * is done with same number of bits as our address mode, so
1850 * we can ignore the conv as address calculation can be viewed
1851 * as either signed or unsigned
1853 set_binop_right(n, pre);
1859 } /* transform_node_AddSub */
1861 #define HANDLE_BINOP_PHI(op,a,b,c) \
1863 if (is_Const(b) && is_const_Phi(a)) { \
1864 /* check for Op(Phi, Const) */ \
1865 c = apply_binop_on_phi(a, get_Const_tarval(b), op, 0); \
1867 else if (is_Const(a) && is_const_Phi(b)) { \
1868 /* check for Op(Const, Phi) */ \
1869 c = apply_binop_on_phi(b, get_Const_tarval(a), op, 1); \
1872 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1876 #define HANDLE_UNOP_PHI(op,a,c) \
1878 if (is_const_Phi(a)) { \
1879 /* check for Op(Phi) */ \
1880 c = apply_unop_on_phi(a, op); \
1883 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1889 * Do the AddSub optimization, then Transform
1890 * Constant folding on Phi
1891 * Add(a,a) -> Mul(a, 2)
1892 * Add(Mul(a, x), a) -> Mul(a, x+1)
1893 * if the mode is integer or float.
1894 * Transform Add(a,-b) into Sub(a,b).
1895 * Reassociation might fold this further.
1897 static ir_node *transform_node_Add(ir_node *n) {
1899 ir_node *a, *b, *c, *oldn = n;
1901 n = transform_node_AddSub(n);
1903 a = get_Add_left(n);
1904 b = get_Add_right(n);
1906 HANDLE_BINOP_PHI(tarval_add, a,b,c);
1908 mode = get_irn_mode(n);
1910 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1911 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1914 if (mode_is_num(mode)) {
1916 ir_node *block = get_irn_n(n, -1);
1919 get_irn_dbg_info(n),
1923 new_r_Const_long(current_ir_graph, block, mode, 2),
1925 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
1926 } else if (get_irn_op(a) == op_Minus) {
1928 get_irn_dbg_info(n),
1934 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1935 } else if (get_irn_op(b) == op_Minus) {
1937 get_irn_dbg_info(n),
1943 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1945 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1946 else if (!get_opt_reassociation() && get_irn_op(a) == op_Mul) {
1947 ir_node *ma = get_Mul_left(a);
1948 ir_node *mb = get_Mul_right(a);
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);
1962 } else if (b == mb) {
1963 ir_node *blk = get_irn_n(n, -1);
1965 get_irn_dbg_info(n), current_ir_graph, blk,
1968 get_irn_dbg_info(n), current_ir_graph, blk,
1970 new_r_Const_long(current_ir_graph, blk, mode, 1),
1973 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1976 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1977 else if (!get_opt_reassociation() && get_irn_op(b) == op_Mul) {
1978 ir_node *ma = get_Mul_left(b);
1979 ir_node *mb = get_Mul_right(b);
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);
1993 } else if (a == mb) {
1994 ir_node *blk = get_irn_n(n, -1);
1996 get_irn_dbg_info(n), current_ir_graph, blk,
1999 get_irn_dbg_info(n), current_ir_graph, blk,
2001 new_r_Const_long(current_ir_graph, blk, mode, 1),
2004 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
2007 /* Here we rely on constants be on the RIGHT side */
2008 else if (get_mode_arithmetic(mode) == irma_twos_complement &&
2009 is_Not(a) && classify_Const(b) == CNST_ONE) {
2011 ir_node *op = get_Not_op(a);
2012 ir_node *blk = get_irn_n(n, -1);
2013 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2014 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2018 } /* transform_node_Add */
2021 * Do the AddSub optimization, then Transform
2022 * Constant folding on Phi
2023 * Sub(0,a) -> Minus(a)
2024 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2025 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2027 static ir_node *transform_node_Sub(ir_node *n) {
2032 n = transform_node_AddSub(n);
2034 a = get_Sub_left(n);
2035 b = get_Sub_right(n);
2037 mode = get_irn_mode(n);
2040 HANDLE_BINOP_PHI(tarval_sub, a,b,c);
2042 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2043 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2047 if (mode_wrap_around(mode)) {
2048 ir_node *left = get_Add_left(a);
2049 ir_node *right = get_Add_right(a);
2051 /* FIXME: Does the Conv's word only for two complement or generally? */
2053 if (mode != get_irn_mode(right)) {
2054 /* This Sub is an effective Cast */
2055 right = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), right, mode);
2058 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2059 } else if (right == b) {
2060 if (mode != get_irn_mode(left)) {
2061 /* This Sub is an effective Cast */
2062 left = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), left, mode);
2065 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2068 } else if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2069 ir_mode *mode = get_irn_mode(a);
2071 if (mode == get_irn_mode(b)) {
2077 /* check if it's allowed to skip the conv */
2078 ma = get_irn_mode(a);
2079 mb = get_irn_mode(b);
2081 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2082 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2084 set_Sub_right(n, b);
2090 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2091 else if (mode_is_num(mode) && mode == get_irn_mode(a) && (classify_Const(a) == CNST_NULL)) {
2093 get_irn_dbg_info(n),
2098 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2100 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2101 else if (get_opt_reassociation() && get_irn_op(a) == op_Mul) {
2102 ir_node *ma = get_Mul_left(a);
2103 ir_node *mb = get_Mul_right(a);
2106 ir_node *blk = get_irn_n(n, -1);
2108 get_irn_dbg_info(n),
2109 current_ir_graph, blk,
2112 get_irn_dbg_info(n),
2113 current_ir_graph, blk,
2115 new_r_Const_long(current_ir_graph, blk, mode, 1),
2118 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2119 } else if (mb == b) {
2120 ir_node *blk = get_irn_n(n, -1);
2122 get_irn_dbg_info(n),
2123 current_ir_graph, blk,
2126 get_irn_dbg_info(n),
2127 current_ir_graph, blk,
2129 new_r_Const_long(current_ir_graph, blk, mode, 1),
2132 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2134 } else if (get_irn_op(a) == op_Sub) {
2135 ir_node *x = get_Sub_left(a);
2136 ir_node *y = get_Sub_right(a);
2137 ir_node *blk = get_irn_n(n, -1);
2138 ir_mode *m_b = get_irn_mode(b);
2139 ir_mode *m_y = get_irn_mode(y);
2142 /* Determine the right mode for the Add. */
2145 else if (mode_is_reference(m_b))
2147 else if (mode_is_reference(m_y))
2151 * Both modes are different but none is reference,
2152 * happens for instance in SubP(SubP(P, Iu), Is).
2153 * We have two possibilities here: Cast or ignore.
2154 * Currently we ignore this case.
2159 add = new_r_Add(current_ir_graph, blk, y, b, mode);
2162 set_Sub_right(n, add);
2163 DBG_OPT_ALGSIM0(n, n, FS_OPT_SUB_SUB_X_Y_Z);
2166 } /* transform_node_Sub */
2169 * Transform Mul(a,-1) into -a.
2170 * Do constant evaluation of Phi nodes.
2171 * Do architecture dependent optimizations on Mul nodes
2173 static ir_node *transform_node_Mul(ir_node *n) {
2174 ir_node *c, *oldn = n;
2175 ir_node *a = get_Mul_left(n);
2176 ir_node *b = get_Mul_right(n);
2179 HANDLE_BINOP_PHI(tarval_mul, a,b,c);
2181 mode = get_irn_mode(n);
2182 if (mode_is_signed(mode)) {
2185 if (value_of(a) == get_mode_minus_one(mode))
2187 else if (value_of(b) == get_mode_minus_one(mode))
2190 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2191 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2195 if (get_mode_arithmetic(mode) == irma_ieee754) {
2197 tarval *tv = get_Const_tarval(a);
2198 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2199 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), b, b, mode);
2200 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2204 else if (is_Const(b)) {
2205 tarval *tv = get_Const_tarval(b);
2206 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2207 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, a, mode);
2208 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2213 return arch_dep_replace_mul_with_shifts(n);
2214 } /* transform_node_Mul */
2217 * Transform a Div Node.
2219 static ir_node *transform_node_Div(ir_node *n) {
2220 tarval *tv = value_of(n);
2221 ir_mode *mode = get_Div_resmode(n);
2224 if (tv != tarval_bad) {
2225 value = new_Const(get_tarval_mode(tv), tv);
2227 DBG_OPT_CSTEVAL(n, value);
2230 ir_node *a = get_Div_left(n);
2231 ir_node *b = get_Div_right(n);
2234 if (a == b && value_not_zero(a, &dummy)) {
2235 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2236 value = new_Const(mode, get_mode_one(mode));
2237 DBG_OPT_CSTEVAL(n, value);
2240 if (mode_is_signed(mode) && is_Const(b)) {
2241 tarval *tv = get_Const_tarval(b);
2243 if (tv == get_mode_minus_one(mode)) {
2245 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2246 DBG_OPT_CSTEVAL(n, value);
2250 /* Try architecture dependent optimization */
2251 value = arch_dep_replace_div_by_const(n);
2259 /* Turn Div into a tuple (mem, jmp, bad, value) */
2260 mem = get_Div_mem(n);
2261 blk = get_irn_n(n, -1);
2263 turn_into_tuple(n, pn_Div_max);
2264 set_Tuple_pred(n, pn_Div_M, mem);
2265 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2266 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2267 set_Tuple_pred(n, pn_Div_res, value);
2270 } /* transform_node_Div */
2273 * Transform a Mod node.
2275 static ir_node *transform_node_Mod(ir_node *n) {
2276 tarval *tv = value_of(n);
2277 ir_mode *mode = get_Mod_resmode(n);
2280 if (tv != tarval_bad) {
2281 value = new_Const(get_tarval_mode(tv), tv);
2283 DBG_OPT_CSTEVAL(n, value);
2286 ir_node *a = get_Mod_left(n);
2287 ir_node *b = get_Mod_right(n);
2290 if (a == b && value_not_zero(a, &dummy)) {
2291 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2292 value = new_Const(mode, get_mode_null(mode));
2293 DBG_OPT_CSTEVAL(n, value);
2296 if (mode_is_signed(mode) && is_Const(b)) {
2297 tarval *tv = get_Const_tarval(b);
2299 if (tv == get_mode_minus_one(mode)) {
2301 value = new_Const(mode, get_mode_null(mode));
2302 DBG_OPT_CSTEVAL(n, value);
2306 /* Try architecture dependent optimization */
2307 value = arch_dep_replace_mod_by_const(n);
2315 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2316 mem = get_Mod_mem(n);
2317 blk = get_irn_n(n, -1);
2319 turn_into_tuple(n, pn_Mod_max);
2320 set_Tuple_pred(n, pn_Mod_M, mem);
2321 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2322 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2323 set_Tuple_pred(n, pn_Mod_res, value);
2326 } /* transform_node_Mod */
2329 * Transform a DivMod node.
2331 static ir_node *transform_node_DivMod(ir_node *n) {
2333 ir_node *a = get_DivMod_left(n);
2334 ir_node *b = get_DivMod_right(n);
2335 ir_mode *mode = get_DivMod_resmode(n);
2336 tarval *ta = value_of(a);
2337 tarval *tb = value_of(b);
2340 if (tb != tarval_bad) {
2341 if (tb == get_mode_one(get_tarval_mode(tb))) {
2342 b = new_Const(mode, get_mode_null(mode));
2343 DBG_OPT_CSTEVAL(n, b);
2345 } else if (ta != tarval_bad) {
2346 tarval *resa, *resb;
2347 resa = tarval_div(ta, tb);
2348 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2349 Jmp for X result!? */
2350 resb = tarval_mod(ta, tb);
2351 if (resb == tarval_bad) return n; /* Causes exception! */
2352 a = new_Const(mode, resa);
2353 b = new_Const(mode, resb);
2354 DBG_OPT_CSTEVAL(n, a);
2355 DBG_OPT_CSTEVAL(n, b);
2357 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
2358 a = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2359 b = new_Const(mode, get_mode_null(mode));
2360 DBG_OPT_CSTEVAL(n, a);
2361 DBG_OPT_CSTEVAL(n, b);
2363 } else { /* Try architecture dependent optimization */
2364 arch_dep_replace_divmod_by_const(&a, &b, n);
2365 evaluated = a != NULL;
2367 } else if (a == b) {
2368 if (value_not_zero(a, &dummy)) {
2370 a = new_Const(mode, get_mode_one(mode));
2371 b = new_Const(mode, get_mode_null(mode));
2372 DBG_OPT_CSTEVAL(n, a);
2373 DBG_OPT_CSTEVAL(n, b);
2376 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2379 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
2380 /* 0 / non-Const = 0 */
2385 if (evaluated) { /* replace by tuple */
2389 mem = get_DivMod_mem(n);
2390 blk = get_irn_n(n, -1);
2391 turn_into_tuple(n, pn_DivMod_max);
2392 set_Tuple_pred(n, pn_DivMod_M, mem);
2393 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
2394 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
2395 set_Tuple_pred(n, pn_DivMod_res_div, a);
2396 set_Tuple_pred(n, pn_DivMod_res_mod, b);
2400 } /* transform_node_DivMod */
2403 * Optimize x / c to x * (1/c)
2405 static ir_node *transform_node_Quot(ir_node *n) {
2406 ir_mode *mode = get_Quot_resmode(n);
2409 if (get_mode_arithmetic(mode) == irma_ieee754) {
2410 ir_node *b = get_Quot_right(n);
2413 tarval *tv = get_Const_tarval(b);
2415 tv = tarval_quo(get_mode_one(mode), tv);
2417 /* Do the transformation if the result is either exact or we are not
2418 using strict rules. */
2419 if (tv != tarval_bad &&
2420 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
2421 ir_node *blk = get_irn_n(n, -1);
2422 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2423 ir_node *a = get_Quot_left(n);
2424 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
2425 ir_node *mem = get_Quot_mem(n);
2427 turn_into_tuple(n, pn_Quot_max);
2428 set_Tuple_pred(n, pn_Quot_M, mem);
2429 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
2430 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
2431 set_Tuple_pred(n, pn_Quot_res, m);
2432 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
2437 } /* transform_node_Quot */
2440 * Optimize Abs(x) into x if x is Confirmed >= 0
2441 * Optimize Abs(x) into -x if x is Confirmed <= 0
2443 static ir_node *transform_node_Abs(ir_node *n) {
2445 ir_node *a = get_Abs_op(n);
2446 value_classify_sign sign = classify_value_sign(a);
2448 if (sign == value_classified_negative) {
2449 ir_mode *mode = get_irn_mode(n);
2452 * We can replace the Abs by -x here.
2453 * We even could add a new Confirm here.
2455 * Note that -x would create a new node, so we could
2456 * not run it in the equivalent_node() context.
2458 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
2459 get_irn_n(n, -1), a, mode);
2461 DBG_OPT_CONFIRM(oldn, n);
2462 } else if (sign == value_classified_positive) {
2463 /* n is positive, Abs is not needed */
2466 DBG_OPT_CONFIRM(oldn, n);
2470 } /* transform_node_Abs */
2473 * Transform a Cond node.
2475 * Replace the Cond by a Jmp if it branches on a constant
2478 static ir_node *transform_node_Cond(ir_node *n) {
2481 ir_node *a = get_Cond_selector(n);
2482 tarval *ta = value_of(a);
2484 /* we need block info which is not available in floating irgs */
2485 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
2488 if ((ta != tarval_bad) &&
2489 (get_irn_mode(a) == mode_b) &&
2490 (get_opt_unreachable_code())) {
2491 /* It's a boolean Cond, branching on a boolean constant.
2492 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2493 jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n));
2494 turn_into_tuple(n, pn_Cond_max);
2495 if (ta == tarval_b_true) {
2496 set_Tuple_pred(n, pn_Cond_false, new_Bad());
2497 set_Tuple_pred(n, pn_Cond_true, jmp);
2499 set_Tuple_pred(n, pn_Cond_false, jmp);
2500 set_Tuple_pred(n, pn_Cond_true, new_Bad());
2502 /* We might generate an endless loop, so keep it alive. */
2503 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
2506 } /* transform_node_Cond */
2508 typedef ir_node* (*recursive_transform) (ir_node *n);
2511 * makes use of distributive laws for and, or, eor
2512 * and(a OP c, b OP c) -> and(a, b) OP c
2514 static ir_node *transform_bitwise_distributive(ir_node *n,
2515 recursive_transform trans_func)
2518 ir_node *a = get_binop_left(n);
2519 ir_node *b = get_binop_right(n);
2520 ir_op *op = get_irn_op(a);
2521 ir_op *op_root = get_irn_op(n);
2523 if(op != get_irn_op(b))
2526 if (op == op_Conv) {
2527 ir_node *a_op = get_Conv_op(a);
2528 ir_node *b_op = get_Conv_op(b);
2529 ir_mode *a_mode = get_irn_mode(a_op);
2530 ir_mode *b_mode = get_irn_mode(b_op);
2531 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
2532 ir_node *blk = get_irn_n(n, -1);
2535 set_binop_left(n, a_op);
2536 set_binop_right(n, b_op);
2537 set_irn_mode(n, a_mode);
2539 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
2541 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
2547 /* nothing to gain here */
2551 if (op == op_Shrs || op == op_Shr || op == op_Shl
2552 || op == op_And || op == op_Or || op == op_Eor) {
2553 ir_node *a_left = get_binop_left(a);
2554 ir_node *a_right = get_binop_right(a);
2555 ir_node *b_left = get_binop_left(b);
2556 ir_node *b_right = get_binop_right(b);
2560 if (is_op_commutative(op)) {
2561 if (a_left == b_left) {
2565 } else if(a_left == b_right) {
2569 } else if(a_right == b_left) {
2575 if(a_right == b_right) {
2582 /* (a sop c) & (b sop c) => (a & b) sop c */
2583 ir_node *blk = get_irn_n(n, -1);
2585 ir_node *new_n = exact_copy(n);
2586 set_binop_left(new_n, op1);
2587 set_binop_right(new_n, op2);
2588 new_n = trans_func(new_n);
2590 if(op_root == op_Eor && op == op_Or) {
2591 dbg_info *dbgi = get_irn_dbg_info(n);
2592 ir_graph *irg = current_ir_graph;
2593 ir_mode *mode = get_irn_mode(c);
2595 c = new_rd_Not(dbgi, irg, blk, c, mode);
2596 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
2599 set_irn_n(n, -1, blk);
2600 set_binop_left(n, new_n);
2601 set_binop_right(n, c);
2605 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
2616 static ir_node *transform_node_And(ir_node *n) {
2618 ir_node *a = get_And_left(n);
2619 ir_node *b = get_And_right(n);
2621 HANDLE_BINOP_PHI(tarval_and, a,b,c);
2623 n = transform_bitwise_distributive(n, transform_node_And);
2626 } /* transform_node_And */
2631 static ir_node *transform_node_Eor(ir_node *n) {
2632 ir_node *c, *oldn = n;
2633 ir_node *a = get_Eor_left(n);
2634 ir_node *b = get_Eor_right(n);
2635 ir_mode *mode = get_irn_mode(n);
2637 HANDLE_BINOP_PHI(tarval_eor, a,b,c);
2641 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
2642 mode, get_mode_null(mode));
2643 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
2644 } else if ((mode == mode_b)
2645 && (get_irn_op(a) == op_Proj)
2646 && (get_irn_mode(a) == mode_b)
2647 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)
2648 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2649 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
2650 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2651 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
2653 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
2654 } else if ((mode == mode_b)
2655 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) {
2656 /* The Eor is a Not. Replace it by a Not. */
2657 /* ????!!!Extend to bitfield 1111111. */
2658 n = new_r_Not(current_ir_graph, get_irn_n(n, -1), a, mode_b);
2660 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2662 n = transform_bitwise_distributive(n, transform_node_Eor);
2666 } /* transform_node_Eor */
2671 static ir_node *transform_node_Not(ir_node *n) {
2672 ir_node *c, *oldn = n;
2673 ir_node *a = get_Not_op(n);
2674 ir_op *op_a = get_irn_op(a);
2676 HANDLE_UNOP_PHI(tarval_not,a,c);
2678 /* check for a boolean Not */
2679 if ( (get_irn_mode(n) == mode_b)
2680 && (op_a == op_Proj)
2681 && (get_irn_mode(a) == mode_b)
2682 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2683 /* We negate a Cmp. The Cmp has the negated result anyways! */
2684 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2685 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
2686 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
2689 if (op_a == op_Sub && classify_Const(get_Sub_right(a)) == CNST_ONE) {
2691 ir_node *op = get_Sub_left(a);
2692 ir_node *blk = get_irn_n(n, -1);
2693 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
2694 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
2697 } /* transform_node_Not */
2700 * Transform a Minus.
2704 static ir_node *transform_node_Minus(ir_node *n) {
2705 ir_node *c, *oldn = n;
2706 ir_node *a = get_Minus_op(n);
2709 HANDLE_UNOP_PHI(tarval_neg,a,c);
2711 mode = get_irn_mode(a);
2712 if (get_mode_arithmetic(mode) == irma_twos_complement && is_Not(a)) {
2714 ir_node *op = get_Not_op(a);
2715 tarval *tv = get_mode_one(mode);
2716 ir_node *blk = get_irn_n(n, -1);
2717 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2718 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
2719 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
2723 } /* transform_node_Minus */
2726 * Transform a Cast_type(Const) into a new Const_type
2728 static ir_node *transform_node_Cast(ir_node *n) {
2730 ir_node *pred = get_Cast_op(n);
2731 ir_type *tp = get_irn_type(n);
2733 if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) {
2734 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
2735 get_Const_tarval(pred), tp);
2736 DBG_OPT_CSTEVAL(oldn, n);
2737 } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) {
2738 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred),
2739 get_SymConst_kind(pred), tp);
2740 DBG_OPT_CSTEVAL(oldn, n);
2744 } /* transform_node_Cast */
2747 * Transform a Proj(Div) with a non-zero value.
2748 * Removes the exceptions and routes the memory to the NoMem node.
2750 static ir_node *transform_node_Proj_Div(ir_node *proj) {
2751 ir_node *div = get_Proj_pred(proj);
2752 ir_node *b = get_Div_right(div);
2753 ir_node *confirm, *res, *new_mem;
2756 if (value_not_zero(b, &confirm)) {
2757 /* div(x, y) && y != 0 */
2758 proj_nr = get_Proj_proj(proj);
2760 case pn_Div_X_regular:
2761 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
2763 case pn_Div_X_except:
2764 /* we found an exception handler, remove it */
2765 DBG_OPT_EXC_REM(proj);
2769 res = get_Div_mem(div);
2770 new_mem = get_irg_no_mem(current_ir_graph);
2773 /* This node can only float up to the Confirm block */
2774 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2776 set_irn_pinned(div, op_pin_state_floats);
2777 /* this is a Div without exception, we can remove the memory edge */
2778 set_Div_mem(div, new_mem);
2783 } /* transform_node_Proj_Div */
2786 * Transform a Proj(Mod) with a non-zero value.
2787 * Removes the exceptions and routes the memory to the NoMem node.
2789 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
2790 ir_node *mod = get_Proj_pred(proj);
2791 ir_node *b = get_Mod_right(mod);
2792 ir_node *confirm, *res, *new_mem;
2795 if (value_not_zero(b, &confirm)) {
2796 /* mod(x, y) && y != 0 */
2797 proj_nr = get_Proj_proj(proj);
2801 case pn_Mod_X_regular:
2802 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
2804 case pn_Mod_X_except:
2805 /* we found an exception handler, remove it */
2806 DBG_OPT_EXC_REM(proj);
2810 res = get_Mod_mem(mod);
2811 new_mem = get_irg_no_mem(current_ir_graph);
2814 /* This node can only float up to the Confirm block */
2815 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2817 set_irn_pinned(mod, op_pin_state_floats);
2818 /* this is a Mod without exception, we can remove the memory edge */
2819 set_Mod_mem(mod, get_irg_no_mem(current_ir_graph));
2822 if (get_Mod_left(mod) == b) {
2823 /* a % a = 0 if a != 0 */
2824 ir_mode *mode = get_irn_mode(proj);
2825 ir_node *res = new_Const(mode, get_mode_null(mode));
2827 DBG_OPT_CSTEVAL(mod, res);
2833 } /* transform_node_Proj_Mod */
2836 * Transform a Proj(DivMod) with a non-zero value.
2837 * Removes the exceptions and routes the memory to the NoMem node.
2839 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
2840 ir_node *divmod = get_Proj_pred(proj);
2841 ir_node *b = get_DivMod_right(divmod);
2842 ir_node *confirm, *res, *new_mem;
2845 if (value_not_zero(b, &confirm)) {
2846 /* DivMod(x, y) && y != 0 */
2847 proj_nr = get_Proj_proj(proj);
2851 case pn_DivMod_X_regular:
2852 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
2854 case pn_DivMod_X_except:
2855 /* we found an exception handler, remove it */
2856 DBG_OPT_EXC_REM(proj);
2860 res = get_DivMod_mem(divmod);
2861 new_mem = get_irg_no_mem(current_ir_graph);
2864 /* This node can only float up to the Confirm block */
2865 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2867 set_irn_pinned(divmod, op_pin_state_floats);
2868 /* this is a DivMod without exception, we can remove the memory edge */
2869 set_DivMod_mem(divmod, get_irg_no_mem(current_ir_graph));
2872 case pn_DivMod_res_mod:
2873 if (get_DivMod_left(divmod) == b) {
2874 /* a % a = 0 if a != 0 */
2875 ir_mode *mode = get_irn_mode(proj);
2876 ir_node *res = new_Const(mode, get_mode_null(mode));
2878 DBG_OPT_CSTEVAL(divmod, res);
2884 } /* transform_node_Proj_DivMod */
2887 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
2889 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
2890 if (get_opt_unreachable_code()) {
2891 ir_node *n = get_Proj_pred(proj);
2892 ir_node *b = get_Cond_selector(n);
2894 if (mode_is_int(get_irn_mode(b))) {
2895 tarval *tb = value_of(b);
2897 if (tb != tarval_bad) {
2898 /* we have a constant switch */
2899 long num = get_Proj_proj(proj);
2901 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
2902 if (get_tarval_long(tb) == num) {
2903 /* Do NOT create a jump here, or we will have 2 control flow ops
2904 * in a block. This case is optimized away in optimize_cf(). */
2907 /* this case will NEVER be taken, kill it */
2915 } /* transform_node_Proj_Cond */
2918 * Normalizes and optimizes Cmp nodes.
2920 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
2921 if (get_opt_reassociation()) {
2922 ir_node *n = get_Proj_pred(proj);
2923 ir_node *left = get_Cmp_left(n);
2924 ir_node *right = get_Cmp_right(n);
2928 ir_mode *mode = NULL;
2929 long proj_nr = get_Proj_proj(proj);
2932 * First step: normalize the compare op
2933 * by placing the constant on the right site
2934 * or moving the lower address node to the left.
2935 * We ignore the case that both are constants
2936 * this case should be optimized away.
2938 if (get_irn_op(right) == op_Const) {
2940 } else if (get_irn_op(left) == op_Const) {
2945 proj_nr = get_inversed_pnc(proj_nr);
2947 } else if (get_irn_idx(left) > get_irn_idx(right)) {
2953 proj_nr = get_inversed_pnc(proj_nr);
2958 * Second step: Try to reduce the magnitude
2959 * of a constant. This may help to generate better code
2960 * later and may help to normalize more compares.
2961 * Of course this is only possible for integer values.
2964 mode = get_irn_mode(c);
2965 tv = get_Const_tarval(c);
2967 if (tv != tarval_bad) {
2968 /* the following optimization is possible on modes without Overflow
2969 * on Unary Minus or on == and !=:
2970 * -a CMP c ==> a swap(CMP) -c
2972 * Beware: for two-complement Overflow may occur, so only == and != can
2973 * be optimized, see this:
2974 * -MININT < 0 =/=> MININT > 0 !!!
2976 if (get_opt_constant_folding() && get_irn_op(left) == op_Minus &&
2977 (!mode_overflow_on_unary_Minus(mode) ||
2978 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
2979 left = get_Minus_op(left);
2980 tv = tarval_neg(tv);
2982 if (tv != tarval_bad) {
2983 proj_nr = get_inversed_pnc(proj_nr);
2988 /* for integer modes, we have more */
2989 if (mode_is_int(mode)) {
2990 /* Ne includes Unordered which is not possible on integers.
2991 * However, frontends often use this wrong, so fix it here */
2992 if (proj_nr & pn_Cmp_Uo) {
2993 proj_nr &= ~pn_Cmp_Uo;
2994 set_Proj_proj(proj, proj_nr);
2997 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
2998 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
2999 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
3000 tv = tarval_sub(tv, get_mode_one(mode));
3002 if (tv != tarval_bad) {
3003 proj_nr ^= pn_Cmp_Eq;
3007 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
3008 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
3009 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
3010 tv = tarval_add(tv, get_mode_one(mode));
3012 if (tv != tarval_bad) {
3013 proj_nr ^= pn_Cmp_Eq;
3018 /* the following reassociations work only for == and != */
3019 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
3021 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
3022 if (classify_tarval(tv) == TV_CLASSIFY_NULL && get_irn_op(left) == op_Sub) {
3023 right = get_Sub_right(left);
3024 left = get_Sub_left(left);
3026 tv = value_of(right);
3027 if (tv != tarval_bad) {
3032 if (tv != tarval_bad) {
3033 ir_op *op = get_irn_op(left);
3035 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
3037 ir_node *c1 = get_Sub_right(left);
3038 tarval *tv2 = value_of(c1);
3040 if (tv2 != tarval_bad) {
3041 tv2 = tarval_add(tv, value_of(c1));
3043 if (tv2 != tarval_bad) {
3044 left = get_Sub_left(left);
3050 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
3051 else if (op == op_Add) {
3052 ir_node *a_l = get_Add_left(left);
3053 ir_node *a_r = get_Add_right(left);
3057 if (get_irn_op(a_l) == op_Const) {
3059 tv2 = value_of(a_l);
3062 tv2 = value_of(a_r);
3065 if (tv2 != tarval_bad) {
3066 tv2 = tarval_sub(tv, tv2);
3068 if (tv2 != tarval_bad) {
3075 /* -a == c ==> a == -c, -a != c ==> a != -c */
3076 else if (op == op_Minus) {
3077 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
3079 if (tv2 != tarval_bad) {
3080 left = get_Minus_op(left);
3087 /* the following reassociations work only for <= */
3088 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3089 if (tv != tarval_bad) {
3090 ir_op *op = get_irn_op(left);
3092 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
3100 * optimization for AND:
3102 * And(x, C) == C ==> And(x, C) != 0
3103 * And(x, C) != C ==> And(X, C) == 0
3105 * if C is a single Bit constant.
3107 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) &&
3108 (get_irn_op(left) == op_And)) {
3109 if (tarval_is_single_bit(tv)) {
3110 /* check for Constant's match. We have check hare the tarvals,
3111 because our const might be changed */
3112 ir_node *la = get_And_left(left);
3113 ir_node *ra = get_And_right(left);
3114 if ((is_Const(la) && get_Const_tarval(la) == tv) ||
3115 (is_Const(ra) && get_Const_tarval(ra) == tv)) {
3116 /* fine: do the transformation */
3117 tv = get_mode_null(get_tarval_mode(tv));
3118 proj_nr ^= pn_Cmp_Leg;
3123 } /* tarval != bad */
3127 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
3129 if (changed & 2) /* need a new Const */
3130 right = new_Const(mode, tv);
3132 /* create a new compare */
3133 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block,
3136 set_Proj_pred(proj, n);
3137 set_Proj_proj(proj, proj_nr);
3141 } /* transform_node_Proj_Cmp */
3144 * Does all optimizations on nodes that must be done on it's Proj's
3145 * because of creating new nodes.
3147 static ir_node *transform_node_Proj(ir_node *proj) {
3148 ir_node *n = get_Proj_pred(proj);
3150 switch (get_irn_opcode(n)) {
3152 return transform_node_Proj_Div(proj);
3155 return transform_node_Proj_Mod(proj);
3158 return transform_node_Proj_DivMod(proj);
3161 return transform_node_Proj_Cond(proj);
3164 return transform_node_Proj_Cmp(proj);
3167 /* should not happen, but if it does will be optimized away */
3168 return equivalent_node_Proj(proj);
3174 } /* transform_node_Proj */
3177 * Move Confirms down through Phi nodes.
3179 static ir_node *transform_node_Phi(ir_node *phi) {
3181 ir_mode *mode = get_irn_mode(phi);
3183 if (mode_is_reference(mode)) {
3184 n = get_irn_arity(phi);
3186 /* Beware of Phi0 */
3188 ir_node *pred = get_irn_n(phi, 0);
3189 ir_node *bound, *new_Phi, *block, **in;
3192 if (! is_Confirm(pred))
3195 bound = get_Confirm_bound(pred);
3196 pnc = get_Confirm_cmp(pred);
3198 NEW_ARR_A(ir_node *, in, n);
3199 in[0] = get_Confirm_value(pred);
3201 for (i = 1; i < n; ++i) {
3202 pred = get_irn_n(phi, i);
3204 if (! is_Confirm(pred) ||
3205 get_Confirm_bound(pred) != bound ||
3206 get_Confirm_cmp(pred) != pnc)
3208 in[i] = get_Confirm_value(pred);
3210 /* move the Confirm nodes "behind" the Phi */
3211 block = get_irn_n(phi, -1);
3212 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
3213 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
3217 } /* transform_node_Phi */
3220 * Returns the operands of a commutative bin-op, if one operand is
3221 * a const, it is returned as the second one.
3223 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
3224 ir_node *op_a = get_binop_left(binop);
3225 ir_node *op_b = get_binop_right(binop);
3227 assert(is_op_commutative(get_irn_op(binop)));
3229 if (get_irn_op(op_a) == op_Const) {
3236 } /* get_comm_Binop_Ops */
3239 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
3240 * Such pattern may arise in bitfield stores.
3242 * value c4 value c4 & c2
3243 * AND c3 AND c1 | c3
3250 * AND c1 ===> OR if (c1 | c2) == 0x111..11
3253 static ir_node *transform_node_Or_bf_store(ir_node *or) {
3256 ir_node *and_l, *c3;
3257 ir_node *value, *c4;
3258 ir_node *new_and, *new_const, *block;
3259 ir_mode *mode = get_irn_mode(or);
3261 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
3264 get_comm_Binop_Ops(or, &and, &c1);
3265 if (!is_Const(c1) || !is_And(and))
3268 get_comm_Binop_Ops(and, &or_l, &c2);
3272 tv1 = get_Const_tarval(c1);
3273 tv2 = get_Const_tarval(c2);
3275 tv = tarval_or(tv1, tv2);
3276 if (classify_tarval(tv) == TV_CLASSIFY_ALL_ONE) {
3277 /* the AND does NOT clear a bit with isn't set be the OR */
3278 set_Or_left(or, or_l);
3279 set_Or_right(or, c1);
3281 /* check for more */
3288 get_comm_Binop_Ops(or_l, &and_l, &c3);
3289 if (!is_Const(c3) || !is_And(and_l))
3292 get_comm_Binop_Ops(and_l, &value, &c4);
3296 /* ok, found the pattern, check for conditions */
3297 assert(mode == get_irn_mode(and));
3298 assert(mode == get_irn_mode(or_l));
3299 assert(mode == get_irn_mode(and_l));
3301 tv3 = get_Const_tarval(c3);
3302 tv4 = get_Const_tarval(c4);
3304 tv = tarval_or(tv4, tv2);
3305 if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) {
3306 /* have at least one 0 at the same bit position */
3310 n_tv4 = tarval_not(tv4);
3311 if (tv3 != tarval_and(tv3, n_tv4)) {
3312 /* bit in the or_mask is outside the and_mask */
3316 n_tv2 = tarval_not(tv2);
3317 if (tv1 != tarval_and(tv1, n_tv2)) {
3318 /* bit in the or_mask is outside the and_mask */
3322 /* ok, all conditions met */
3323 block = get_irn_n(or, -1);
3325 new_and = new_r_And(current_ir_graph, block,
3326 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
3328 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
3330 set_Or_left(or, new_and);
3331 set_Or_right(or, new_const);
3333 /* check for more */
3335 } /* transform_node_Or_bf_store */
3338 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
3340 static ir_node *transform_node_Or_Rot(ir_node *or) {
3341 ir_mode *mode = get_irn_mode(or);
3342 ir_node *shl, *shr, *block;
3343 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
3346 if (! mode_is_int(mode))
3349 shl = get_binop_left(or);
3350 shr = get_binop_right(or);
3352 if (get_irn_op(shl) == op_Shr) {
3353 if (get_irn_op(shr) != op_Shl)
3359 } else if (get_irn_op(shl) != op_Shl) {
3361 } else if (get_irn_op(shr) != op_Shr) {
3364 x = get_Shl_left(shl);
3365 if (x != get_Shr_left(shr))
3368 c1 = get_Shl_right(shl);
3369 c2 = get_Shr_right(shr);
3370 if (get_irn_op(c1) == op_Const && get_irn_op(c2) == op_Const) {
3371 tv1 = get_Const_tarval(c1);
3372 if (! tarval_is_long(tv1))
3375 tv2 = get_Const_tarval(c2);
3376 if (! tarval_is_long(tv2))
3379 if (get_tarval_long(tv1) + get_tarval_long(tv2)
3380 != get_mode_size_bits(mode))
3383 /* yet, condition met */
3384 block = get_irn_n(or, -1);
3386 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
3388 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
3390 } else if (get_irn_op(c1) == op_Sub) {
3394 if (get_Sub_right(sub) != v)
3397 c1 = get_Sub_left(sub);
3398 if (get_irn_op(c1) != op_Const)
3401 tv1 = get_Const_tarval(c1);
3402 if (! tarval_is_long(tv1))
3405 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
3408 /* yet, condition met */
3409 block = get_nodes_block(or);
3411 /* a Rot right is not supported, so use a rot left */
3412 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
3414 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
3416 } else if (get_irn_op(c2) == op_Sub) {
3420 c1 = get_Sub_left(sub);
3421 if (get_irn_op(c1) != op_Const)
3424 tv1 = get_Const_tarval(c1);
3425 if (! tarval_is_long(tv1))
3428 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
3431 /* yet, condition met */
3432 block = get_irn_n(or, -1);
3435 n = new_r_Rot(current_ir_graph, block, x, v, mode);
3437 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
3442 } /* transform_node_Or_Rot */
3447 static ir_node *transform_node_Or(ir_node *n) {
3448 ir_node *c, *oldn = n;
3449 ir_node *a = get_Or_left(n);
3450 ir_node *b = get_Or_right(n);
3452 HANDLE_BINOP_PHI(tarval_or, a,b,c);
3454 n = transform_node_Or_bf_store(n);
3455 n = transform_node_Or_Rot(n);
3459 n = transform_bitwise_distributive(n, transform_node_Or);
3462 } /* transform_node_Or */
3466 static ir_node *transform_node(ir_node *n);
3469 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl.
3471 * Should be moved to reassociation?
3473 static ir_node *transform_node_shift(ir_node *n) {
3474 ir_node *left, *right;
3475 tarval *tv1, *tv2, *res;
3477 int modulo_shf, flag;
3479 left = get_binop_left(n);
3481 /* different operations */
3482 if (get_irn_op(left) != get_irn_op(n))
3485 right = get_binop_right(n);
3486 tv1 = value_of(right);
3487 if (tv1 == tarval_bad)
3490 tv2 = value_of(get_binop_right(left));
3491 if (tv2 == tarval_bad)
3494 res = tarval_add(tv1, tv2);
3496 /* beware: a simple replacement works only, if res < modulo shift */
3497 mode = get_irn_mode(n);
3501 modulo_shf = get_mode_modulo_shift(mode);
3502 if (modulo_shf > 0) {
3503 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
3505 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
3511 /* ok, we can replace it */
3512 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
3514 in[0] = get_binop_left(left);
3515 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
3517 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
3519 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
3521 return transform_node(irn);
3524 } /* transform_node_shift */
3529 static ir_node *transform_node_Shr(ir_node *n) {
3530 ir_node *c, *oldn = n;
3531 ir_node *a = get_Shr_left(n);
3532 ir_node *b = get_Shr_right(n);
3534 HANDLE_BINOP_PHI(tarval_shr, a, b, c);
3535 return transform_node_shift(n);
3536 } /* transform_node_Shr */
3541 static ir_node *transform_node_Shrs(ir_node *n) {
3542 ir_node *c, *oldn = n;
3543 ir_node *a = get_Shrs_left(n);
3544 ir_node *b = get_Shrs_right(n);
3546 HANDLE_BINOP_PHI(tarval_shrs, a, b, c);
3547 return transform_node_shift(n);
3548 } /* transform_node_Shrs */
3553 static ir_node *transform_node_Shl(ir_node *n) {
3554 ir_node *c, *oldn = n;
3555 ir_node *a = get_Shl_left(n);
3556 ir_node *b = get_Shl_right(n);
3558 HANDLE_BINOP_PHI(tarval_shl, a, b, c);
3559 return transform_node_shift(n);
3560 } /* transform_node_Shl */
3563 * Remove dead blocks and nodes in dead blocks
3564 * in keep alive list. We do not generate a new End node.
3566 static ir_node *transform_node_End(ir_node *n) {
3567 int i, j, n_keepalives = get_End_n_keepalives(n);
3570 NEW_ARR_A(ir_node *, in, n_keepalives);
3572 for (i = j = 0; i < n_keepalives; ++i) {
3573 ir_node *ka = get_End_keepalive(n, i);
3575 if (! is_Block_dead(ka)) {
3579 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
3582 /* FIXME: beabi need to keep a Proj(M) */
3583 if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka))
3586 if (j != n_keepalives)
3587 set_End_keepalives(n, j, in);
3589 } /* transform_node_End */
3592 * Optimize a Mux into some simpler cases.
3594 static ir_node *transform_node_Mux(ir_node *n) {
3595 ir_node *oldn = n, *sel = get_Mux_sel(n);
3596 ir_mode *mode = get_irn_mode(n);
3598 if (mode == mode_b) {
3599 ir_node *t = get_Mux_true(n);
3600 ir_node *f = get_Mux_false(n);
3601 dbg_info *dbg = get_irn_dbg_info(n);
3602 ir_node *block = get_irn_n(n, -1);
3603 ir_graph *irg = current_ir_graph;
3606 tarval *tv_t = get_Const_tarval(t);
3607 if (tv_t == tarval_b_true) {
3609 assert(get_Const_tarval(f) == tarval_b_false);
3612 return new_rd_Or(dbg, irg, block, sel, f, mode_b);
3615 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
3616 assert(tv_t == tarval_b_false);
3618 assert(get_Const_tarval(f) == tarval_b_true);
3621 return new_rd_And(dbg, irg, block, not_sel, f, mode_b);
3624 } else if (is_Const(f)) {
3625 tarval *tv_f = get_Const_tarval(f);
3626 if (tv_f == tarval_b_true) {
3627 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
3628 return new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
3630 assert(tv_f == tarval_b_false);
3631 return new_rd_And(dbg, irg, block, sel, t, mode_b);
3636 if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(mode)) {
3637 ir_node *cmp = get_Proj_pred(sel);
3638 long proj_nr = get_Proj_proj(sel);
3639 ir_node *f = get_Mux_false(n);
3640 ir_node *t = get_Mux_true(n);
3642 if (get_irn_op(cmp) == op_Cmp && classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
3643 ir_node *block = get_irn_n(n, -1);
3646 * Note: normalization puts the constant on the right site,
3647 * so we check only one case.
3649 * Note further that these optimization work even for floating point
3650 * with NaN's because -NaN == NaN.
3651 * However, if +0 and -0 is handled differently, we cannot use the first one.
3653 if (get_irn_op(f) == op_Minus &&
3654 get_Minus_op(f) == t &&
3655 get_Cmp_left(cmp) == t) {
3657 if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3658 /* Mux(a >=/> 0, -a, a) ==> Abs(a) */
3659 n = new_rd_Abs(get_irn_dbg_info(n),
3663 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3665 } else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3666 /* Mux(a <=/< 0, -a, a) ==> Minus(Abs(a)) */
3667 n = new_rd_Abs(get_irn_dbg_info(n),
3671 n = new_rd_Minus(get_irn_dbg_info(n),
3676 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3679 } else if (get_irn_op(t) == op_Minus &&
3680 get_Minus_op(t) == f &&
3681 get_Cmp_left(cmp) == f) {
3683 if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3684 /* Mux(a <=/< 0, a, -a) ==> Abs(a) */
3685 n = new_rd_Abs(get_irn_dbg_info(n),
3689 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3691 } else if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3692 /* Mux(a >=/> 0, a, -a) ==> Minus(Abs(a)) */
3693 n = new_rd_Abs(get_irn_dbg_info(n),
3697 n = new_rd_Minus(get_irn_dbg_info(n),
3702 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3707 if (mode_is_int(mode) && mode_is_signed(mode) &&
3708 get_mode_arithmetic(mode) == irma_twos_complement) {
3709 ir_node *x = get_Cmp_left(cmp);
3711 /* the following optimization works only with signed integer two-complement mode */
3713 if (mode == get_irn_mode(x)) {
3715 * FIXME: this restriction is two rigid, as it would still
3716 * work if mode(x) = Hs and mode == Is, but at least it removes
3719 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Le) &&
3720 classify_Const(t) == CNST_ALL_ONE &&
3721 classify_Const(f) == CNST_NULL) {
3723 * Mux(x:T </<= 0, 0, -1) -> Shrs(x, sizeof_bits(T) - 1)
3727 n = new_rd_Shrs(get_irn_dbg_info(n),
3728 current_ir_graph, block, x,
3729 new_r_Const_long(current_ir_graph, block, mode_Iu,
3730 get_mode_size_bits(mode) - 1),
3732 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3734 } else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) &&
3735 classify_Const(t) == CNST_ONE &&
3736 classify_Const(f) == CNST_NULL) {
3738 * Mux(x:T >/>= 0, 0, 1) -> Shr(-x, sizeof_bits(T) - 1)
3742 n = new_rd_Shr(get_irn_dbg_info(n),
3743 current_ir_graph, block,
3744 new_r_Minus(current_ir_graph, block, x, mode),
3745 new_r_Const_long(current_ir_graph, block, mode_Iu,
3746 get_mode_size_bits(mode) - 1),
3748 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3755 return arch_transform_node_Mux(n);
3756 } /* transform_node_Mux */
3759 * Optimize a Psi into some simpler cases.
3761 static ir_node *transform_node_Psi(ir_node *n) {
3763 return transform_node_Mux(n);
3766 } /* transform_node_Psi */
3769 * Tries several [inplace] [optimizing] transformations and returns an
3770 * equivalent node. The difference to equivalent_node() is that these
3771 * transformations _do_ generate new nodes, and thus the old node must
3772 * not be freed even if the equivalent node isn't the old one.
3774 static ir_node *transform_node(ir_node *n) {
3778 * Transform_node is the only "optimizing transformation" that might
3779 * return a node with a different opcode. We iterate HERE until fixpoint
3780 * to get the final result.
3784 if (n->op->ops.transform_node)
3785 n = n->op->ops.transform_node(n);
3786 } while (oldn != n);
3789 } /* transform_node */
3792 * Sets the default transform node operation for an ir_op_ops.
3794 * @param code the opcode for the default operation
3795 * @param ops the operations initialized
3800 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
3804 ops->transform_node = transform_node_##a; \
3838 } /* firm_set_default_transform_node */
3841 /* **************** Common Subexpression Elimination **************** */
3843 /** The size of the hash table used, should estimate the number of nodes
3845 #define N_IR_NODES 512
3847 /** Compares the attributes of two Const nodes. */
3848 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
3849 return (get_Const_tarval(a) != get_Const_tarval(b))
3850 || (get_Const_type(a) != get_Const_type(b));
3851 } /* node_cmp_attr_Const */
3853 /** Compares the attributes of two Proj nodes. */
3854 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
3855 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
3856 } /* node_cmp_attr_Proj */
3858 /** Compares the attributes of two Filter nodes. */
3859 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
3860 return get_Filter_proj(a) != get_Filter_proj(b);
3861 } /* node_cmp_attr_Filter */
3863 /** Compares the attributes of two Alloc nodes. */
3864 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
3865 const alloc_attr *pa = get_irn_alloc_attr(a);
3866 const alloc_attr *pb = get_irn_alloc_attr(b);
3867 return (pa->where != pb->where) || (pa->type != pb->type);
3868 } /* node_cmp_attr_Alloc */
3870 /** Compares the attributes of two Free nodes. */
3871 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
3872 const free_attr *pa = get_irn_free_attr(a);
3873 const free_attr *pb = get_irn_free_attr(b);
3874 return (pa->where != pb->where) || (pa->type != pb->type);
3875 } /* node_cmp_attr_Free */
3877 /** Compares the attributes of two SymConst nodes. */
3878 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
3879 const symconst_attr *pa = get_irn_symconst_attr(a);
3880 const symconst_attr *pb = get_irn_symconst_attr(b);
3881 return (pa->num != pb->num)
3882 || (pa->sym.type_p != pb->sym.type_p)
3883 || (pa->tp != pb->tp);
3884 } /* node_cmp_attr_SymConst */
3886 /** Compares the attributes of two Call nodes. */
3887 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
3888 return (get_irn_call_attr(a) != get_irn_call_attr(b));
3889 } /* node_cmp_attr_Call */
3891 /** Compares the attributes of two Sel nodes. */
3892 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
3893 const ir_entity *a_ent = get_Sel_entity(a);
3894 const ir_entity *b_ent = get_Sel_entity(b);
3896 (a_ent->kind != b_ent->kind) ||
3897 (a_ent->name != b_ent->name) ||
3898 (a_ent->owner != b_ent->owner) ||
3899 (a_ent->ld_name != b_ent->ld_name) ||
3900 (a_ent->type != b_ent->type);
3901 } /* node_cmp_attr_Sel */
3903 /** Compares the attributes of two Phi nodes. */
3904 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
3905 /* we can only enter this function if both nodes have the same number of inputs,
3906 hence it is enough to check if one of them is a Phi0 */
3908 /* check the Phi0 attribute */
3909 return get_irn_phi0_attr(a) != get_irn_phi0_attr(b);
3912 } /* node_cmp_attr_Phi */
3914 /** Compares the attributes of two Conv nodes. */
3915 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
3916 return get_Conv_strict(a) != get_Conv_strict(b);
3917 } /* node_cmp_attr_Conv */
3919 /** Compares the attributes of two Cast nodes. */
3920 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
3921 return get_Cast_type(a) != get_Cast_type(b);
3922 } /* node_cmp_attr_Cast */
3924 /** Compares the attributes of two Load nodes. */
3925 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
3926 if (get_Load_volatility(a) == volatility_is_volatile ||
3927 get_Load_volatility(b) == volatility_is_volatile)
3928 /* NEVER do CSE on volatile Loads */
3930 /* do not CSE Loads with different alignment. Be conservative. */
3931 if (get_Load_align(a) != get_Load_align(b))
3934 return get_Load_mode(a) != get_Load_mode(b);
3935 } /* node_cmp_attr_Load */
3937 /** Compares the attributes of two Store nodes. */
3938 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
3939 /* do not CSE Stores with different alignment. Be conservative. */
3940 if (get_Store_align(a) != get_Store_align(b))
3943 /* NEVER do CSE on volatile Stores */
3944 return (get_Store_volatility(a) == volatility_is_volatile ||
3945 get_Store_volatility(b) == volatility_is_volatile);
3946 } /* node_cmp_attr_Store */
3948 /** Compares the attributes of two Confirm nodes. */
3949 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
3950 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
3951 } /* node_cmp_attr_Confirm */
3953 /** Compares the attributes of two ASM nodes. */
3954 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
3956 const ir_asm_constraint *ca;
3957 const ir_asm_constraint *cb;
3960 if (get_ASM_text(a) != get_ASM_text(b))
3963 /* Should we really check the constraints here? Should be better, but is strange. */
3964 n = get_ASM_n_input_constraints(a);
3965 if (n != get_ASM_n_input_constraints(b))
3968 ca = get_ASM_input_constraints(a);
3969 cb = get_ASM_input_constraints(b);
3970 for (i = 0; i < n; ++i) {
3971 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
3975 n = get_ASM_n_output_constraints(a);
3976 if (n != get_ASM_n_output_constraints(b))
3979 ca = get_ASM_output_constraints(a);
3980 cb = get_ASM_output_constraints(b);
3981 for (i = 0; i < n; ++i) {
3982 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
3986 n = get_ASM_n_clobbers(a);
3987 if (n != get_ASM_n_clobbers(b))
3990 cla = get_ASM_clobbers(a);
3991 clb = get_ASM_clobbers(b);
3992 for (i = 0; i < n; ++i) {
3993 if (cla[i] != clb[i])
3997 } /* node_cmp_attr_ASM */
4000 * Set the default node attribute compare operation for an ir_op_ops.
4002 * @param code the opcode for the default operation
4003 * @param ops the operations initialized
4008 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
4012 ops->node_cmp_attr = node_cmp_attr_##a; \
4037 } /* firm_set_default_node_cmp_attr */
4040 * Compare function for two nodes in the hash table. Gets two
4041 * nodes as parameters. Returns 0 if the nodes are a cse.
4043 int identities_cmp(const void *elt, const void *key) {
4050 if (a == b) return 0;
4052 if ((get_irn_op(a) != get_irn_op(b)) ||
4053 (get_irn_mode(a) != get_irn_mode(b))) return 1;
4055 /* compare if a's in and b's in are of equal length */
4056 irn_arity_a = get_irn_intra_arity (a);
4057 if (irn_arity_a != get_irn_intra_arity(b))
4060 /* for block-local cse and op_pin_state_pinned nodes: */
4061 if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) {
4062 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
4066 /* compare a->in[0..ins] with b->in[0..ins] */
4067 for (i = 0; i < irn_arity_a; i++)
4068 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
4072 * here, we already now that the nodes are identical except their
4075 if (a->op->ops.node_cmp_attr)
4076 return a->op->ops.node_cmp_attr(a, b);
4079 } /* identities_cmp */
4082 * Calculate a hash value of a node.
4084 unsigned ir_node_hash(ir_node *node) {
4088 if (node->op == op_Const) {
4089 /* special value for const, as they only differ in their tarval. */
4090 h = HASH_PTR(node->attr.con.tv);
4091 h = 9*h + HASH_PTR(get_irn_mode(node));
4092 } else if (node->op == op_SymConst) {
4093 /* special value for const, as they only differ in their symbol. */
4094 h = HASH_PTR(node->attr.symc.sym.type_p);
4095 h = 9*h + HASH_PTR(get_irn_mode(node));
4098 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
4099 h = irn_arity = get_irn_intra_arity(node);
4101 /* consider all in nodes... except the block if not a control flow. */
4102 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
4103 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
4107 h = 9*h + HASH_PTR(get_irn_mode(node));
4109 h = 9*h + HASH_PTR(get_irn_op(node));
4113 } /* ir_node_hash */
4115 pset *new_identities(void) {
4116 return new_pset(identities_cmp, N_IR_NODES);
4117 } /* new_identities */
4119 void del_identities(pset *value_table) {
4120 del_pset(value_table);
4121 } /* del_identities */
4124 * Return the canonical node computing the same value as n.
4126 * @param value_table The value table
4127 * @param n The node to lookup
4129 * Looks up the node in a hash table.
4131 * For Const nodes this is performed in the constructor, too. Const
4132 * nodes are extremely time critical because of their frequent use in
4133 * constant string arrays.
4135 static INLINE ir_node *identify(pset *value_table, ir_node *n) {
4138 if (!value_table) return n;
4140 if (get_opt_reassociation()) {
4141 if (is_op_commutative(get_irn_op(n))) {
4142 ir_node *l = get_binop_left(n);
4143 ir_node *r = get_binop_right(n);
4145 /* for commutative operators perform a OP b == b OP a */
4146 if (get_irn_idx(l) > get_irn_idx(r)) {
4147 set_binop_left(n, r);
4148 set_binop_right(n, l);
4153 o = pset_find(value_table, n, ir_node_hash(n));
4162 * During construction we set the op_pin_state_pinned flag in the graph right when the
4163 * optimization is performed. The flag turning on procedure global cse could
4164 * be changed between two allocations. This way we are safe.
4166 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
4169 n = identify(value_table, n);
4170 if (get_irn_n(old, -1) != get_irn_n(n, -1))
4171 set_irg_pinned(current_ir_graph, op_pin_state_floats);
4173 } /* identify_cons */
4176 * Return the canonical node computing the same value as n.
4177 * Looks up the node in a hash table, enters it in the table
4178 * if it isn't there yet.
4180 ir_node *identify_remember(pset *value_table, ir_node *n) {
4183 if (!value_table) return n;
4185 if (get_opt_reassociation()) {
4186 if (is_op_commutative(get_irn_op(n))) {
4187 ir_node *l = get_binop_left(n);
4188 ir_node *r = get_binop_right(n);
4189 int l_idx = get_irn_idx(l);
4190 int r_idx = get_irn_idx(r);
4192 /* For commutative operators perform a OP b == b OP a but keep
4193 constants on the RIGHT side. This helps greatly in some optimizations.
4194 Moreover we use the idx number to make the form deterministic. */
4195 if (is_irn_constlike(l))
4197 if (is_irn_constlike(r))
4199 if (l_idx < r_idx) {
4200 set_binop_left(n, r);
4201 set_binop_right(n, l);
4206 /* lookup or insert in hash table with given hash key. */
4207 o = pset_insert(value_table, n, ir_node_hash(n));
4214 } /* identify_remember */
4216 /* Add a node to the identities value table. */
4217 void add_identities(pset *value_table, ir_node *node) {
4218 if (get_opt_cse() && is_no_Block(node))
4219 identify_remember(value_table, node);
4220 } /* add_identities */
4222 /* Visit each node in the value table of a graph. */
4223 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
4225 ir_graph *rem = current_ir_graph;
4227 current_ir_graph = irg;
4228 foreach_pset(irg->value_table, node)
4230 current_ir_graph = rem;
4231 } /* visit_all_identities */
4234 * Garbage in, garbage out. If a node has a dead input, i.e., the
4235 * Bad node is input to the node, return the Bad node.
4237 static INLINE ir_node *gigo(ir_node *node) {
4239 ir_op *op = get_irn_op(node);
4241 /* remove garbage blocks by looking at control flow that leaves the block
4242 and replacing the control flow by Bad. */
4243 if (get_irn_mode(node) == mode_X) {
4244 ir_node *block = get_nodes_block(skip_Proj(node));
4246 /* Don't optimize nodes in immature blocks. */
4247 if (!get_Block_matured(block)) return node;
4248 /* Don't optimize End, may have Bads. */
4249 if (op == op_End) return node;
4251 if (is_Block(block)) {
4252 irn_arity = get_irn_arity(block);
4253 for (i = 0; i < irn_arity; i++) {
4254 if (!is_Bad(get_irn_n(block, i)))
4257 if (i == irn_arity) {
4258 ir_graph *irg = get_irn_irg(block);
4259 /* the start block is never dead */
4260 if (block != get_irg_start_block(irg)
4261 && block != get_irg_end_block(irg))
4267 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
4268 blocks predecessors is dead. */
4269 if (op != op_Block && op != op_Phi && op != op_Tuple) {
4270 irn_arity = get_irn_arity(node);
4273 * Beware: we can only read the block of a non-floating node.
4275 if (is_irn_pinned_in_irg(node) &&
4276 is_Block_dead(get_nodes_block(node)))
4279 for (i = 0; i < irn_arity; i++) {
4280 ir_node *pred = get_irn_n(node, i);
4285 /* Propagating Unknowns here seems to be a bad idea, because
4286 sometimes we need a node as a input and did not want that
4288 However, it might be useful to move this into a later phase
4289 (if you think that optimizing such code is useful). */
4290 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
4291 return new_Unknown(get_irn_mode(node));
4296 /* With this code we violate the agreement that local_optimize
4297 only leaves Bads in Block, Phi and Tuple nodes. */
4298 /* If Block has only Bads as predecessors it's garbage. */
4299 /* If Phi has only Bads as predecessors it's garbage. */
4300 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
4301 irn_arity = get_irn_arity(node);
4302 for (i = 0; i < irn_arity; i++) {
4303 if (!is_Bad(get_irn_n(node, i))) break;
4305 if (i == irn_arity) node = new_Bad();
4312 * These optimizations deallocate nodes from the obstack.
4313 * It can only be called if it is guaranteed that no other nodes
4314 * reference this one, i.e., right after construction of a node.
4316 * @param n The node to optimize
4318 * current_ir_graph must be set to the graph of the node!
4320 ir_node *optimize_node(ir_node *n) {
4323 ir_opcode iro = get_irn_opcode(n);
4325 /* Always optimize Phi nodes: part of the construction. */
4326 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
4328 /* constant expression evaluation / constant folding */
4329 if (get_opt_constant_folding()) {
4330 /* neither constants nor Tuple values can be evaluated */
4331 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
4332 /* try to evaluate */
4333 tv = computed_value(n);
4334 if (tv != tarval_bad) {
4336 ir_type *old_tp = get_irn_type(n);
4337 int i, arity = get_irn_arity(n);
4341 * Try to recover the type of the new expression.
4343 for (i = 0; i < arity && !old_tp; ++i)
4344 old_tp = get_irn_type(get_irn_n(n, i));
4347 * we MUST copy the node here temporary, because it's still needed
4348 * for DBG_OPT_CSTEVAL
4350 node_size = offsetof(ir_node, attr) + n->op->attr_size;
4351 oldn = alloca(node_size);
4353 memcpy(oldn, n, node_size);
4354 CLONE_ARR_A(ir_node *, oldn->in, n->in);
4356 /* ARG, copy the in array, we need it for statistics */
4357 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
4359 /* note the inplace edges module */
4360 edges_node_deleted(n, current_ir_graph);
4362 /* evaluation was successful -- replace the node. */
4363 irg_kill_node(current_ir_graph, n);
4364 nw = new_Const(get_tarval_mode (tv), tv);
4366 if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv))
4367 set_Const_type(nw, old_tp);
4368 DBG_OPT_CSTEVAL(oldn, nw);
4374 /* remove unnecessary nodes */
4375 if (get_opt_constant_folding() ||
4376 (iro == iro_Phi) || /* always optimize these nodes. */
4378 (iro == iro_Proj) ||
4379 (iro == iro_Block) ) /* Flags tested local. */
4380 n = equivalent_node(n);
4382 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
4384 /* Common Subexpression Elimination.
4386 * Checks whether n is already available.
4387 * The block input is used to distinguish different subexpressions. Right
4388 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
4389 * subexpressions within a block.
4392 n = identify_cons(current_ir_graph->value_table, n);
4395 edges_node_deleted(oldn, current_ir_graph);
4397 /* We found an existing, better node, so we can deallocate the old node. */
4398 irg_kill_node(current_ir_graph, oldn);
4402 /* Some more constant expression evaluation that does not allow to
4404 iro = get_irn_opcode(n);
4405 if (get_opt_constant_folding() ||
4406 (iro == iro_Cond) ||
4407 (iro == iro_Proj)) /* Flags tested local. */
4408 n = transform_node(n);
4410 /* Remove nodes with dead (Bad) input.
4411 Run always for transformation induced Bads. */
4414 /* Now we have a legal, useful node. Enter it in hash table for CSE */
4415 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
4416 n = identify_remember(current_ir_graph->value_table, n);
4420 } /* optimize_node */
4424 * These optimizations never deallocate nodes (in place). This can cause dead
4425 * nodes lying on the obstack. Remove these by a dead node elimination,
4426 * i.e., a copying garbage collection.
4428 ir_node *optimize_in_place_2(ir_node *n) {
4431 ir_opcode iro = get_irn_opcode(n);
4433 if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n;
4435 /* constant expression evaluation / constant folding */
4436 if (get_opt_constant_folding()) {
4437 /* neither constants nor Tuple values can be evaluated */
4438 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
4439 /* try to evaluate */
4440 tv = computed_value(n);
4441 if (tv != tarval_bad) {
4442 /* evaluation was successful -- replace the node. */
4443 ir_type *old_tp = get_irn_type(n);
4444 int i, arity = get_irn_arity(n);
4447 * Try to recover the type of the new expression.
4449 for (i = 0; i < arity && !old_tp; ++i)
4450 old_tp = get_irn_type(get_irn_n(n, i));
4452 n = new_Const(get_tarval_mode(tv), tv);
4454 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
4455 set_Const_type(n, old_tp);
4457 DBG_OPT_CSTEVAL(oldn, n);
4463 /* remove unnecessary nodes */
4464 if (get_opt_constant_folding() ||
4465 (iro == iro_Phi) || /* always optimize these nodes. */
4466 (iro == iro_Id) || /* ... */
4467 (iro == iro_Proj) || /* ... */
4468 (iro == iro_Block) ) /* Flags tested local. */
4469 n = equivalent_node(n);
4471 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
4473 /** common subexpression elimination **/
4474 /* Checks whether n is already available. */
4475 /* The block input is used to distinguish different subexpressions. Right
4476 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
4477 subexpressions within a block. */
4478 if (get_opt_cse()) {
4479 n = identify(current_ir_graph->value_table, n);
4482 /* Some more constant expression evaluation. */
4483 iro = get_irn_opcode(n);
4484 if (get_opt_constant_folding() ||
4485 (iro == iro_Cond) ||
4486 (iro == iro_Proj)) /* Flags tested local. */
4487 n = transform_node(n);
4489 /* Remove nodes with dead (Bad) input.
4490 Run always for transformation induced Bads. */
4493 /* Now we can verify the node, as it has no dead inputs any more. */
4496 /* Now we have a legal, useful node. Enter it in hash table for cse.
4497 Blocks should be unique anyways. (Except the successor of start:
4498 is cse with the start block!) */
4499 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
4500 n = identify_remember(current_ir_graph->value_table, n);
4503 } /* optimize_in_place_2 */
4506 * Wrapper for external use, set proper status bits after optimization.
4508 ir_node *optimize_in_place(ir_node *n) {
4509 /* Handle graph state */
4510 assert(get_irg_phase_state(current_ir_graph) != phase_building);
4512 if (get_opt_global_cse())
4513 set_irg_pinned(current_ir_graph, op_pin_state_floats);
4514 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
4515 set_irg_outs_inconsistent(current_ir_graph);
4517 /* FIXME: Maybe we could also test whether optimizing the node can
4518 change the control graph. */
4519 set_irg_doms_inconsistent(current_ir_graph);
4520 return optimize_in_place_2(n);
4521 } /* optimize_in_place */
4524 * Sets the default operation for an ir_ops.
4526 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
4527 ops = firm_set_default_computed_value(code, ops);
4528 ops = firm_set_default_equivalent_node(code, ops);
4529 ops = firm_set_default_transform_node(code, ops);
4530 ops = firm_set_default_node_cmp_attr(code, ops);
4531 ops = firm_set_default_get_type(code, ops);
4532 ops = firm_set_default_get_type_attr(code, ops);
4533 ops = firm_set_default_get_entity_attr(code, ops);
4536 } /* firm_set_default_operations */