2 * Copyright (C) 1995-2008 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 * Returns the tarval of a Const node or tarval_bad for all other nodes.
59 static tarval *default_value_of(const ir_node *n) {
61 return get_Const_tarval(n); /* might return tarval_bad */
66 value_of_func value_of_ptr = default_value_of;
68 /* * Set a new value_of function. */
69 void set_value_of_func(value_of_func func) {
73 value_of_ptr = default_value_of;
77 * Return the value of a Constant.
79 static tarval *computed_value_Const(const ir_node *n) {
80 return get_Const_tarval(n);
81 } /* computed_value_Const */
84 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
86 static tarval *computed_value_SymConst(const ir_node *n) {
90 switch (get_SymConst_kind(n)) {
91 case symconst_type_size:
92 type = get_SymConst_type(n);
93 if (get_type_state(type) == layout_fixed)
94 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
96 case symconst_type_align:
97 type = get_SymConst_type(n);
98 if (get_type_state(type) == layout_fixed)
99 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
101 case symconst_ofs_ent:
102 ent = get_SymConst_entity(n);
103 type = get_entity_owner(ent);
104 if (get_type_state(type) == layout_fixed)
105 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
111 } /* computed_value_SymConst */
114 * Return the value of an Add.
116 static tarval *computed_value_Add(const ir_node *n) {
117 ir_node *a = get_Add_left(n);
118 ir_node *b = get_Add_right(n);
120 tarval *ta = value_of(a);
121 tarval *tb = value_of(b);
123 if ((ta != tarval_bad) && (tb != tarval_bad))
124 return tarval_add(ta, tb);
127 } /* computed_value_Add */
130 * Return the value of a Sub.
131 * Special case: a - a
133 static tarval *computed_value_Sub(const ir_node *n) {
134 ir_mode *mode = get_irn_mode(n);
135 ir_node *a = get_Sub_left(n);
136 ir_node *b = get_Sub_right(n);
141 if (a == b && !is_Bad(a))
142 return get_mode_null(mode);
147 if ((ta != tarval_bad) && (tb != tarval_bad))
148 return tarval_sub(ta, tb, mode);
151 } /* computed_value_Sub */
154 * Return the value of a Carry.
155 * Special : a op 0, 0 op b
157 static tarval *computed_value_Carry(const ir_node *n) {
158 ir_node *a = get_binop_left(n);
159 ir_node *b = get_binop_right(n);
160 ir_mode *m = get_irn_mode(n);
162 tarval *ta = value_of(a);
163 tarval *tb = value_of(b);
165 if ((ta != tarval_bad) && (tb != tarval_bad)) {
167 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
169 if (tarval_is_null(ta) || tarval_is_null(tb))
170 return get_mode_null(m);
173 } /* computed_value_Carry */
176 * Return the value of a Borrow.
179 static tarval *computed_value_Borrow(const ir_node *n) {
180 ir_node *a = get_binop_left(n);
181 ir_node *b = get_binop_right(n);
182 ir_mode *m = get_irn_mode(n);
184 tarval *ta = value_of(a);
185 tarval *tb = value_of(b);
187 if ((ta != tarval_bad) && (tb != tarval_bad)) {
188 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
189 } else if (tarval_is_null(ta)) {
190 return get_mode_null(m);
193 } /* computed_value_Borrow */
196 * Return the value of an unary Minus.
198 static tarval *computed_value_Minus(const ir_node *n) {
199 ir_node *a = get_Minus_op(n);
200 tarval *ta = value_of(a);
202 if (ta != tarval_bad)
203 return tarval_neg(ta);
206 } /* computed_value_Minus */
209 * Return the value of a Mul.
211 static tarval *computed_value_Mul(const ir_node *n) {
212 ir_node *a = get_Mul_left(n);
213 ir_node *b = get_Mul_right(n);
216 tarval *ta = value_of(a);
217 tarval *tb = value_of(b);
219 mode = get_irn_mode(n);
220 if (mode != get_irn_mode(a)) {
221 /* n * n = 2n bit multiplication */
222 ta = tarval_convert_to(ta, mode);
223 tb = tarval_convert_to(tb, mode);
226 if (ta != tarval_bad && tb != tarval_bad) {
227 return tarval_mul(ta, tb);
229 /* a*0 = 0 or 0*b = 0 */
230 if (ta == get_mode_null(mode))
232 if (tb == get_mode_null(mode))
236 } /* computed_value_Mul */
239 * Return the value of an Abs.
241 static tarval *computed_value_Abs(const ir_node *n) {
242 ir_node *a = get_Abs_op(n);
243 tarval *ta = value_of(a);
245 if (ta != tarval_bad)
246 return tarval_abs(ta);
249 } /* computed_value_Abs */
252 * Return the value of an And.
253 * Special case: a & 0, 0 & b
255 static tarval *computed_value_And(const ir_node *n) {
256 ir_node *a = get_And_left(n);
257 ir_node *b = get_And_right(n);
259 tarval *ta = value_of(a);
260 tarval *tb = value_of(b);
262 if ((ta != tarval_bad) && (tb != tarval_bad)) {
263 return tarval_and (ta, tb);
265 if (tarval_is_null(ta)) return ta;
266 if (tarval_is_null(tb)) return tb;
269 } /* computed_value_And */
272 * Return the value of an Or.
273 * Special case: a | 1...1, 1...1 | b
275 static tarval *computed_value_Or(const ir_node *n) {
276 ir_node *a = get_Or_left(n);
277 ir_node *b = get_Or_right(n);
279 tarval *ta = value_of(a);
280 tarval *tb = value_of(b);
282 if ((ta != tarval_bad) && (tb != tarval_bad)) {
283 return tarval_or (ta, tb);
285 if (tarval_is_all_one(ta)) return ta;
286 if (tarval_is_all_one(tb)) return tb;
289 } /* computed_value_Or */
292 * Return the value of an Eor.
294 static tarval *computed_value_Eor(const ir_node *n) {
295 ir_node *a = get_Eor_left(n);
296 ir_node *b = get_Eor_right(n);
301 return get_mode_null(get_irn_mode(n));
306 if ((ta != tarval_bad) && (tb != tarval_bad)) {
307 return tarval_eor (ta, tb);
310 } /* computed_value_Eor */
313 * Return the value of a Not.
315 static tarval *computed_value_Not(const ir_node *n) {
316 ir_node *a = get_Not_op(n);
317 tarval *ta = value_of(a);
319 if (ta != tarval_bad)
320 return tarval_not(ta);
323 } /* computed_value_Not */
326 * Return the value of a Shl.
328 static tarval *computed_value_Shl(const ir_node *n) {
329 ir_node *a = get_Shl_left(n);
330 ir_node *b = get_Shl_right(n);
332 tarval *ta = value_of(a);
333 tarval *tb = value_of(b);
335 if ((ta != tarval_bad) && (tb != tarval_bad)) {
336 return tarval_shl (ta, tb);
339 } /* computed_value_Shl */
342 * Return the value of a Shr.
344 static tarval *computed_value_Shr(const ir_node *n) {
345 ir_node *a = get_Shr_left(n);
346 ir_node *b = get_Shr_right(n);
348 tarval *ta = value_of(a);
349 tarval *tb = value_of(b);
351 if ((ta != tarval_bad) && (tb != tarval_bad)) {
352 return tarval_shr (ta, tb);
355 } /* computed_value_Shr */
358 * Return the value of a Shrs.
360 static tarval *computed_value_Shrs(const ir_node *n) {
361 ir_node *a = get_Shrs_left(n);
362 ir_node *b = get_Shrs_right(n);
364 tarval *ta = value_of(a);
365 tarval *tb = value_of(b);
367 if ((ta != tarval_bad) && (tb != tarval_bad)) {
368 return tarval_shrs (ta, tb);
371 } /* computed_value_Shrs */
374 * Return the value of a Rotl.
376 static tarval *computed_value_Rotl(const ir_node *n) {
377 ir_node *a = get_Rotl_left(n);
378 ir_node *b = get_Rotl_right(n);
380 tarval *ta = value_of(a);
381 tarval *tb = value_of(b);
383 if ((ta != tarval_bad) && (tb != tarval_bad)) {
384 return tarval_rotl(ta, tb);
387 } /* computed_value_Rotl */
390 * Return the value of a Conv.
392 static tarval *computed_value_Conv(const ir_node *n) {
393 ir_node *a = get_Conv_op(n);
394 tarval *ta = value_of(a);
396 if (ta != tarval_bad)
397 return tarval_convert_to(ta, get_irn_mode(n));
400 } /* computed_value_Conv */
403 * Calculate the value of a Mux: can be evaluated, if the
404 * sel and the right input are known.
406 static tarval *computed_value_Mux(const ir_node *n) {
407 ir_node *sel = get_Mux_sel(n);
408 tarval *ts = value_of(sel);
410 if (ts == get_tarval_b_true()) {
411 ir_node *v = get_Mux_true(n);
414 else if (ts == get_tarval_b_false()) {
415 ir_node *v = get_Mux_false(n);
419 } /* computed_value_Mux */
422 * Calculate the value of a Psi: can be evaluated, if a condition is true
423 * and all previous conditions are false. If all conditions are false
424 * we evaluate to the default one.
426 static tarval *computed_value_Psi(const ir_node *n) {
428 return computed_value_Mux(n);
430 } /* computed_value_Psi */
433 * Calculate the value of a Confirm: can be evaluated,
434 * if it has the form Confirm(x, '=', Const).
436 static tarval *computed_value_Confirm(const ir_node *n) {
438 * Beware: we might produce Phi(Confirm(x == true), Confirm(x == false)).
439 * Do NOT optimize them away (CondEval wants them), so wait until
440 * remove_confirm is activated.
442 if (get_opt_remove_confirm()) {
443 if (get_Confirm_cmp(n) == pn_Cmp_Eq) {
444 tarval *tv = value_of(get_Confirm_bound(n));
445 if (tv != tarval_bad)
449 return value_of(get_Confirm_value(n));
450 } /* computed_value_Confirm */
453 * Return the value of a Proj(Cmp).
455 * This performs a first step of unreachable code elimination.
456 * Proj can not be computed, but folding a Cmp above the Proj here is
457 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
459 * There are several case where we can evaluate a Cmp node, see later.
461 static tarval *computed_value_Proj_Cmp(const ir_node *n) {
462 ir_node *a = get_Proj_pred(n);
463 ir_node *aa = get_Cmp_left(a);
464 ir_node *ab = get_Cmp_right(a);
465 long proj_nr = get_Proj_proj(n);
468 * BEWARE: a == a is NOT always True for floating Point values, as
469 * NaN != NaN is defined, so we must check this here.
472 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
475 /* This is a trick with the bits used for encoding the Cmp
476 Proj numbers, the following statement is not the same:
477 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
478 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
481 tarval *taa = value_of(aa);
482 tarval *tab = value_of(ab);
483 ir_mode *mode = get_irn_mode(aa);
486 * The predecessors of Cmp are target values. We can evaluate
489 if ((taa != tarval_bad) && (tab != tarval_bad)) {
490 /* strange checks... */
491 pn_Cmp flags = tarval_cmp(taa, tab);
492 if (flags != pn_Cmp_False) {
493 return new_tarval_from_long (proj_nr & flags, mode_b);
496 /* for integer values, we can check against MIN/MAX */
497 else if (mode_is_int(mode)) {
498 /* MIN <=/> x. This results in true/false. */
499 if (taa == get_mode_min(mode)) {
500 /* a compare with the MIN value */
501 if (proj_nr == pn_Cmp_Le)
502 return get_tarval_b_true();
503 else if (proj_nr == pn_Cmp_Gt)
504 return get_tarval_b_false();
506 /* x >=/< MIN. This results in true/false. */
508 if (tab == get_mode_min(mode)) {
509 /* a compare with the MIN value */
510 if (proj_nr == pn_Cmp_Ge)
511 return get_tarval_b_true();
512 else if (proj_nr == pn_Cmp_Lt)
513 return get_tarval_b_false();
515 /* MAX >=/< x. This results in true/false. */
516 else if (taa == get_mode_max(mode)) {
517 if (proj_nr == pn_Cmp_Ge)
518 return get_tarval_b_true();
519 else if (proj_nr == pn_Cmp_Lt)
520 return get_tarval_b_false();
522 /* x <=/> MAX. This results in true/false. */
523 else if (tab == get_mode_max(mode)) {
524 if (proj_nr == pn_Cmp_Le)
525 return get_tarval_b_true();
526 else if (proj_nr == pn_Cmp_Gt)
527 return get_tarval_b_false();
531 * The predecessors are Allocs or (void*)(0) constants. Allocs never
532 * return NULL, they raise an exception. Therefore we can predict
536 ir_node *aaa = skip_Id(skip_Proj(aa));
537 ir_node *aba = skip_Id(skip_Proj(ab));
539 if ( ( (/* aa is ProjP and aaa is Alloc */
541 && mode_is_reference(get_irn_mode(aa))
543 && ( (/* ab is NULL */
545 && mode_is_reference(get_irn_mode(ab))
546 && is_Const_null(ab))
547 || (/* ab is other Alloc */
549 && mode_is_reference(get_irn_mode(ab))
552 || (/* aa is NULL and aba is Alloc */
554 && mode_is_reference(get_irn_mode(aa))
557 && mode_is_reference(get_irn_mode(ab))
560 return new_tarval_from_long(proj_nr & pn_Cmp_Ne, mode_b);
563 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
564 } /* computed_value_Proj_Cmp */
567 * Return the value of a floating point Quot.
569 static tarval *do_computed_value_Quot(const ir_node *a, const ir_node *b) {
570 tarval *ta = value_of(a);
571 tarval *tb = value_of(b);
573 /* cannot optimize 0 / b = 0 because of NaN */
574 if (ta != tarval_bad && tb != tarval_bad)
575 return tarval_quo(ta, tb);
577 } /* do_computed_value_Quot */
580 * Calculate the value of an integer Div of two nodes.
581 * Special case: 0 / b
583 static tarval *do_computed_value_Div(const ir_node *a, const ir_node *b) {
584 tarval *ta = value_of(a);
586 const ir_node *dummy;
588 /* Compute c1 / c2 or 0 / a, a != 0 */
589 if (tarval_is_null(ta) && value_not_zero(b, &dummy))
590 return ta; /* 0 / b == 0 */
592 if (ta != tarval_bad && tb != tarval_bad)
593 return tarval_div(ta, tb);
595 } /* do_computed_value_Div */
598 * Calculate the value of an integer Mod of two nodes.
599 * Special case: a % 1
601 static tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b) {
602 tarval *ta = value_of(a);
603 tarval *tb = value_of(b);
605 /* Compute a % 1 or c1 % c2 */
606 if (tarval_is_one(tb))
607 return get_mode_null(get_irn_mode(a));
608 if (ta != tarval_bad && tb != tarval_bad)
609 return tarval_mod(ta, tb);
611 } /* do_computed_value_Mod */
614 * Return the value of a Proj(DivMod).
616 static tarval *computed_value_Proj_DivMod(const ir_node *n) {
617 long proj_nr = get_Proj_proj(n);
619 /* compute either the Div or the Mod part */
620 if (proj_nr == pn_DivMod_res_div) {
621 const ir_node *a = get_Proj_pred(n);
622 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
623 } else if (proj_nr == pn_DivMod_res_mod) {
624 const ir_node *a = get_Proj_pred(n);
625 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
628 } /* computed_value_Proj_DivMod */
631 * Return the value of a Proj(Div).
633 static tarval *computed_value_Proj_Div(const ir_node *n) {
634 long proj_nr = get_Proj_proj(n);
636 if (proj_nr == pn_Div_res) {
637 const ir_node *a = get_Proj_pred(n);
638 return do_computed_value_Div(get_Div_left(a), get_Div_right(a));
641 } /* computed_value_Proj_Div */
644 * Return the value of a Proj(Mod).
646 static tarval *computed_value_Proj_Mod(const ir_node *n) {
647 long proj_nr = get_Proj_proj(n);
649 if (proj_nr == pn_Mod_res) {
650 const ir_node *a = get_Proj_pred(n);
651 return do_computed_value_Mod(get_Mod_left(a), get_Mod_right(a));
654 } /* computed_value_Proj_Mod */
657 * Return the value of a Proj(Quot).
659 static tarval *computed_value_Proj_Quot(const ir_node *n) {
660 long proj_nr = get_Proj_proj(n);
662 if (proj_nr == pn_Quot_res) {
663 const ir_node *a = get_Proj_pred(n);
664 return do_computed_value_Quot(get_Quot_left(a), get_Quot_right(a));
667 } /* computed_value_Proj_Quot */
670 * Return the value of a Proj.
672 static tarval *computed_value_Proj(const ir_node *proj) {
673 ir_node *n = get_Proj_pred(proj);
675 if (n->op->ops.computed_value_Proj != NULL)
676 return n->op->ops.computed_value_Proj(proj);
678 } /* computed_value_Proj */
681 * If the parameter n can be computed, return its value, else tarval_bad.
682 * Performs constant folding.
684 * @param n The node this should be evaluated
686 tarval *computed_value(const ir_node *n) {
687 if (n->op->ops.computed_value)
688 return n->op->ops.computed_value(n);
690 } /* computed_value */
693 * Set the default computed_value evaluator in an ir_op_ops.
695 * @param code the opcode for the default operation
696 * @param ops the operations initialized
701 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
705 ops->computed_value = computed_value_##a; \
707 #define CASE_PROJ(a) \
709 ops->computed_value_Proj = computed_value_Proj_##a; \
747 } /* firm_set_default_computed_value */
750 * Returns a equivalent block for another block.
751 * If the block has only one predecessor, this is
752 * the equivalent one. If the only predecessor of a block is
753 * the block itself, this is a dead block.
755 * If both predecessors of a block are the branches of a binary
756 * Cond, the equivalent block is Cond's block.
758 * If all predecessors of a block are bad or lies in a dead
759 * block, the current block is dead as well.
761 * Note, that blocks are NEVER turned into Bad's, instead
762 * the dead_block flag is set. So, never test for is_Bad(block),
763 * always use is_dead_Block(block).
765 static ir_node *equivalent_node_Block(ir_node *n)
770 /* don't optimize dead blocks */
771 if (is_Block_dead(n))
774 n_preds = get_Block_n_cfgpreds(n);
776 /* The Block constructor does not call optimize, but mature_immBlock()
777 calls the optimization. */
778 assert(get_Block_matured(n));
780 /* Straightening: a single entry Block following a single exit Block
781 can be merged, if it is not the Start block. */
782 /* !!! Beware, all Phi-nodes of n must have been optimized away.
783 This should be true, as the block is matured before optimize is called.
784 But what about Phi-cycles with the Phi0/Id that could not be resolved?
785 Remaining Phi nodes are just Ids. */
787 ir_node *pred = skip_Proj(get_Block_cfgpred(n, 0));
790 ir_node *predblock = get_nodes_block(pred);
791 if (predblock == oldn) {
792 /* Jmp jumps into the block it is in -- deal self cycle. */
793 n = set_Block_dead(n);
794 DBG_OPT_DEAD_BLOCK(oldn, n);
795 } else if (get_opt_control_flow_straightening()) {
797 DBG_OPT_STG(oldn, n);
799 } else if (is_Cond(pred)) {
800 ir_node *predblock = get_nodes_block(pred);
801 if (predblock == oldn) {
802 /* Jmp jumps into the block it is in -- deal self cycle. */
803 n = set_Block_dead(n);
804 DBG_OPT_DEAD_BLOCK(oldn, n);
807 } else if ((n_preds == 2) &&
808 (get_opt_control_flow_weak_simplification())) {
809 /* Test whether Cond jumps twice to this block
810 * The more general case which more than 2 predecessors is handles
811 * in optimize_cf(), we handle only this special case for speed here.
813 ir_node *a = get_Block_cfgpred(n, 0);
814 ir_node *b = get_Block_cfgpred(n, 1);
816 if (is_Proj(a) && is_Proj(b)) {
817 ir_node *cond = get_Proj_pred(a);
819 if (cond == get_Proj_pred(b) && is_Cond(cond) &&
820 get_irn_mode(get_Cond_selector(cond)) == mode_b) {
821 /* Also a single entry Block following a single exit Block. Phis have
822 twice the same operand and will be optimized away. */
823 n = get_nodes_block(cond);
824 DBG_OPT_IFSIM1(oldn, a, b, n);
827 } else if (get_opt_unreachable_code() &&
828 (n != get_irg_start_block(current_ir_graph)) &&
829 (n != get_irg_end_block(current_ir_graph)) ) {
832 /* If all inputs are dead, this block is dead too, except if it is
833 the start or end block. This is one step of unreachable code
835 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
836 ir_node *pred = get_Block_cfgpred(n, i);
839 if (is_Bad(pred)) continue;
840 pred_blk = get_nodes_block(skip_Proj(pred));
842 if (is_Block_dead(pred_blk)) continue;
845 /* really found a living input */
850 n = set_Block_dead(n);
851 DBG_OPT_DEAD_BLOCK(oldn, n);
856 } /* equivalent_node_Block */
859 * Returns a equivalent node for a Jmp, a Bad :-)
860 * Of course this only happens if the Block of the Jmp is dead.
862 static ir_node *equivalent_node_Jmp(ir_node *n) {
865 /* unreachable code elimination */
866 if (is_Block_dead(get_nodes_block(n))) {
867 n = get_irg_bad(current_ir_graph);
868 DBG_OPT_DEAD_BLOCK(oldn, n);
871 } /* equivalent_node_Jmp */
873 /** Raise is handled in the same way as Jmp. */
874 #define equivalent_node_Raise equivalent_node_Jmp
877 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
878 See transform_node_Proj_Cond(). */
881 * Optimize operations that are commutative and have neutral 0,
882 * so a op 0 = 0 op a = a.
884 static ir_node *equivalent_node_neutral_zero(ir_node *n) {
887 ir_node *a = get_binop_left(n);
888 ir_node *b = get_binop_right(n);
893 /* After running compute_node there is only one constant predecessor.
894 Find this predecessors value and remember the other node: */
895 if ((tv = value_of(a)) != tarval_bad) {
897 } else if ((tv = value_of(b)) != tarval_bad) {
902 /* If this predecessors constant value is zero, the operation is
903 * unnecessary. Remove it.
905 * Beware: If n is a Add, the mode of on and n might be different
906 * which happens in this rare construction: NULL + 3.
907 * Then, a Conv would be needed which we cannot include here.
909 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
912 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
916 } /* equivalent_node_neutral_zero */
919 * Eor is commutative and has neutral 0.
921 static ir_node *equivalent_node_Eor(ir_node *n) {
926 n = equivalent_node_neutral_zero(n);
927 if (n != oldn) return n;
930 b = get_Eor_right(n);
933 ir_node *aa = get_Eor_left(a);
934 ir_node *ab = get_Eor_right(a);
937 /* (a ^ b) ^ a -> b */
939 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
941 } else if (ab == b) {
942 /* (a ^ b) ^ b -> a */
944 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
949 ir_node *ba = get_Eor_left(b);
950 ir_node *bb = get_Eor_right(b);
953 /* a ^ (a ^ b) -> b */
955 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
957 } else if (bb == a) {
958 /* a ^ (b ^ a) -> b */
960 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
968 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
970 * The second one looks strange, but this construct
971 * is used heavily in the LCC sources :-).
973 * Beware: The Mode of an Add may be different than the mode of its
974 * predecessors, so we could not return a predecessors in all cases.
976 static ir_node *equivalent_node_Add(ir_node *n) {
978 ir_node *left, *right;
979 ir_mode *mode = get_irn_mode(n);
981 n = equivalent_node_neutral_zero(n);
985 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
986 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
989 left = get_Add_left(n);
990 right = get_Add_right(n);
993 if (get_Sub_right(left) == right) {
996 n = get_Sub_left(left);
997 if (mode == get_irn_mode(n)) {
998 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
1003 if (is_Sub(right)) {
1004 if (get_Sub_right(right) == left) {
1007 n = get_Sub_left(right);
1008 if (mode == get_irn_mode(n)) {
1009 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
1015 } /* equivalent_node_Add */
1018 * optimize operations that are not commutative but have neutral 0 on left,
1021 static ir_node *equivalent_node_left_zero(ir_node *n) {
1024 ir_node *a = get_binop_left(n);
1025 ir_node *b = get_binop_right(n);
1026 tarval *tb = value_of(b);
1028 if (tarval_is_null(tb)) {
1031 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1034 } /* equivalent_node_left_zero */
1036 #define equivalent_node_Shl equivalent_node_left_zero
1037 #define equivalent_node_Shr equivalent_node_left_zero
1038 #define equivalent_node_Shrs equivalent_node_left_zero
1039 #define equivalent_node_Rotl equivalent_node_left_zero
1042 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
1044 * The second one looks strange, but this construct
1045 * is used heavily in the LCC sources :-).
1047 * Beware: The Mode of a Sub may be different than the mode of its
1048 * predecessors, so we could not return a predecessors in all cases.
1050 static ir_node *equivalent_node_Sub(ir_node *n) {
1053 ir_mode *mode = get_irn_mode(n);
1056 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1057 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1060 b = get_Sub_right(n);
1063 /* Beware: modes might be different */
1064 if (tarval_is_null(tb)) {
1065 ir_node *a = get_Sub_left(n);
1066 if (mode == get_irn_mode(a)) {
1069 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1073 } /* equivalent_node_Sub */
1077 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
1080 * -(-a) == a, but might overflow two times.
1081 * We handle it anyway here but the better way would be a
1082 * flag. This would be needed for Pascal for instance.
1084 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
1086 ir_node *pred = get_unop_op(n);
1088 /* optimize symmetric unop */
1089 if (get_irn_op(pred) == get_irn_op(n)) {
1090 n = get_unop_op(pred);
1091 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
1094 } /* equivalent_node_idempotent_unop */
1096 /** Optimize Not(Not(x)) == x. */
1097 #define equivalent_node_Not equivalent_node_idempotent_unop
1099 /** -(-x) == x ??? Is this possible or can --x raise an
1100 out of bounds exception if min =! max? */
1101 #define equivalent_node_Minus equivalent_node_idempotent_unop
1104 * Optimize a * 1 = 1 * a = a.
1106 static ir_node *equivalent_node_Mul(ir_node *n) {
1108 ir_node *a = get_Mul_left(n);
1110 /* we can handle here only the n * n = n bit cases */
1111 if (get_irn_mode(n) == get_irn_mode(a)) {
1112 ir_node *b = get_Mul_right(n);
1116 * Mul is commutative and has again an other neutral element.
1117 * Constants are place right, so check this case first.
1120 if (tarval_is_one(tv)) {
1122 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1125 if (tarval_is_one(tv)) {
1127 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1132 } /* equivalent_node_Mul */
1135 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1137 static ir_node *equivalent_node_Or(ir_node *n) {
1140 ir_node *a = get_Or_left(n);
1141 ir_node *b = get_Or_right(n);
1145 n = a; /* Or has it's own neutral element */
1146 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1149 /* constants are cormalized to right, check this site first */
1151 if (tarval_is_null(tv)) {
1153 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1157 if (tarval_is_null(tv)) {
1159 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1164 } /* equivalent_node_Or */
1167 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1169 static ir_node *equivalent_node_And(ir_node *n) {
1172 ir_node *a = get_And_left(n);
1173 ir_node *b = get_And_right(n);
1177 n = a; /* And has it's own neutral element */
1178 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1181 /* constants are cormalized to right, check this site first */
1183 if (tarval_is_all_one(tv)) {
1185 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1189 if (tarval_is_all_one(tv)) {
1191 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1195 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1198 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1203 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1206 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1211 } /* equivalent_node_And */
1214 * Try to remove useless Conv's:
1216 static ir_node *equivalent_node_Conv(ir_node *n) {
1218 ir_node *a = get_Conv_op(n);
1220 ir_mode *n_mode = get_irn_mode(n);
1221 ir_mode *a_mode = get_irn_mode(a);
1224 if (n_mode == a_mode) { /* No Conv necessary */
1225 if (get_Conv_strict(n)) {
1226 /* special case: the predecessor might be a also a Conv */
1228 if (! get_Conv_strict(a)) {
1229 /* first one is not strict, kick it */
1231 a_mode = get_irn_mode(a);
1235 /* else both are strict conv, second is superfluous */
1236 } else if (is_Proj(a)) {
1237 ir_node *pred = get_Proj_pred(a);
1238 if (is_Load(pred)) {
1239 /* loads always return with the exact precision of n_mode */
1240 assert(get_Load_mode(pred) == n_mode);
1245 /* leave strict floating point Conv's */
1249 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1250 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1251 ir_node *b = get_Conv_op(a);
1252 ir_mode *b_mode = get_irn_mode(b);
1254 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1255 /* both are strict conv */
1256 if (smaller_mode(a_mode, n_mode)) {
1257 /* both are strict, but the first is smaller, so
1258 the second cannot remove more precision, remove the
1260 set_Conv_strict(n, 0);
1263 if (n_mode == b_mode) {
1264 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1265 if (n_mode == mode_b) {
1266 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1267 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1268 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1269 if (smaller_mode(b_mode, a_mode)) {
1270 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1271 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1275 if (mode_is_int(n_mode) && mode_is_float(a_mode)) {
1276 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1277 size_t int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1278 size_t float_mantissa;
1279 /* FIXME There is no way to get the mantissa size of a mode */
1280 switch (get_mode_size_bits(a_mode)) {
1281 case 32: float_mantissa = 23 + 1; break; // + 1 for implicit 1
1282 case 64: float_mantissa = 52 + 1; break;
1283 case 80: float_mantissa = 64 + 1; break;
1284 default: float_mantissa = 0; break;
1286 if (float_mantissa != 0 && float_mantissa >= int_mantissa) {
1288 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1293 if (smaller_mode(b_mode, a_mode)) {
1294 if (get_Conv_strict(n))
1295 set_Conv_strict(b, 1);
1296 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1297 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1303 } /* equivalent_node_Conv */
1306 * A Cast may be removed if the type of the previous node
1307 * is already the type of the Cast.
1309 static ir_node *equivalent_node_Cast(ir_node *n) {
1311 ir_node *pred = get_Cast_op(n);
1313 if (get_irn_type(pred) == get_Cast_type(n)) {
1315 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1318 } /* equivalent_node_Cast */
1321 * - fold Phi-nodes, iff they have only one predecessor except
1324 static ir_node *equivalent_node_Phi(ir_node *n) {
1329 ir_node *first_val = NULL; /* to shutup gcc */
1331 if (!get_opt_normalize()) return n;
1333 n_preds = get_Phi_n_preds(n);
1335 block = get_nodes_block(n);
1336 if (is_Block_dead(block)) /* Control dead */
1337 return get_irg_bad(current_ir_graph);
1339 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1341 /* Find first non-self-referencing input */
1342 for (i = 0; i < n_preds; ++i) {
1343 first_val = get_Phi_pred(n, i);
1344 if ( (first_val != n) /* not self pointer */
1346 /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad
1347 * predecessors. Then, Phi nodes in dead code might be removed, causing
1348 * nodes pointing to themself (Add's for instance).
1349 * This is really bad and causes endless recursions in several
1350 * code pathes, so we do NOT optimize such a code.
1351 * This is not that bad as it sounds, optimize_cf() removes bad control flow
1352 * (and bad Phi predecessors), so live code is optimized later.
1354 && (! is_Bad(get_Block_cfgpred(block, i)))
1356 ) { /* value not dead */
1357 break; /* then found first value. */
1362 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1363 return get_irg_bad(current_ir_graph);
1366 /* search for rest of inputs, determine if any of these
1367 are non-self-referencing */
1368 while (++i < n_preds) {
1369 ir_node *scnd_val = get_Phi_pred(n, i);
1370 if ( (scnd_val != n)
1371 && (scnd_val != first_val)
1374 && (! is_Bad(get_Block_cfgpred(block, i)))
1382 /* Fold, if no multiple distinct non-self-referencing inputs */
1384 DBG_OPT_PHI(oldn, n);
1387 } /* equivalent_node_Phi */
1390 * Several optimizations:
1391 * - fold Sync-nodes, iff they have only one predecessor except
1394 static ir_node *equivalent_node_Sync(ir_node *n) {
1395 int arity = get_Sync_n_preds(n);
1398 for (i = 0; i < arity;) {
1399 ir_node *pred = get_Sync_pred(n, i);
1402 /* Remove Bad predecessors */
1409 /* Remove duplicate predecessors */
1415 if (get_Sync_pred(n, j) == pred) {
1423 if (arity == 0) return get_irg_bad(current_ir_graph);
1424 if (arity == 1) return get_Sync_pred(n, 0);
1426 } /* equivalent_node_Sync */
1429 * Optimize Proj(Tuple).
1431 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj) {
1432 ir_node *oldn = proj;
1433 ir_node *tuple = get_Proj_pred(proj);
1435 /* Remove the Tuple/Proj combination. */
1436 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1437 DBG_OPT_TUPLE(oldn, tuple, proj);
1440 } /* equivalent_node_Proj_Tuple */
1443 * Optimize a / 1 = a.
1445 static ir_node *equivalent_node_Proj_Div(ir_node *proj) {
1446 ir_node *oldn = proj;
1447 ir_node *div = get_Proj_pred(proj);
1448 ir_node *b = get_Div_right(div);
1449 tarval *tb = value_of(b);
1451 /* Div is not commutative. */
1452 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1453 switch (get_Proj_proj(proj)) {
1455 proj = get_Div_mem(div);
1456 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1460 proj = get_Div_left(div);
1461 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1465 /* we cannot replace the exception Proj's here, this is done in
1466 transform_node_Proj_Div() */
1471 } /* equivalent_node_Proj_Div */
1474 * Optimize a / 1.0 = a.
1476 static ir_node *equivalent_node_Proj_Quot(ir_node *proj) {
1477 ir_node *oldn = proj;
1478 ir_node *quot = get_Proj_pred(proj);
1479 ir_node *b = get_Quot_right(quot);
1480 tarval *tb = value_of(b);
1482 /* Div is not commutative. */
1483 if (tarval_is_one(tb)) { /* Quot(x, 1) == x */
1484 switch (get_Proj_proj(proj)) {
1486 proj = get_Quot_mem(quot);
1487 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1491 proj = get_Quot_left(quot);
1492 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1496 /* we cannot replace the exception Proj's here, this is done in
1497 transform_node_Proj_Quot() */
1502 } /* equivalent_node_Proj_Quot */
1505 * Optimize a / 1 = a.
1507 static ir_node *equivalent_node_Proj_DivMod(ir_node *proj) {
1508 ir_node *oldn = proj;
1509 ir_node *divmod = get_Proj_pred(proj);
1510 ir_node *b = get_DivMod_right(divmod);
1511 tarval *tb = value_of(b);
1513 /* Div is not commutative. */
1514 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1515 switch (get_Proj_proj(proj)) {
1517 proj = get_DivMod_mem(divmod);
1518 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1521 case pn_DivMod_res_div:
1522 proj = get_DivMod_left(divmod);
1523 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1527 /* we cannot replace the exception Proj's here, this is done in
1528 transform_node_Proj_DivMod().
1529 Note further that the pn_DivMod_res_div case is handled in
1530 computed_value_Proj(). */
1535 } /* equivalent_node_Proj_DivMod */
1538 * Optimize CopyB(mem, x, x) into a Nop.
1540 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj) {
1541 ir_node *oldn = proj;
1542 ir_node *copyb = get_Proj_pred(proj);
1543 ir_node *a = get_CopyB_dst(copyb);
1544 ir_node *b = get_CopyB_src(copyb);
1547 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1548 switch (get_Proj_proj(proj)) {
1549 case pn_CopyB_M_regular:
1550 proj = get_CopyB_mem(copyb);
1551 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1554 case pn_CopyB_M_except:
1555 case pn_CopyB_X_except:
1556 DBG_OPT_EXC_REM(proj);
1557 proj = get_irg_bad(current_ir_graph);
1562 } /* equivalent_node_Proj_CopyB */
1565 * Optimize Bounds(idx, idx, upper) into idx.
1567 static ir_node *equivalent_node_Proj_Bound(ir_node *proj) {
1568 ir_node *oldn = proj;
1569 ir_node *bound = get_Proj_pred(proj);
1570 ir_node *idx = get_Bound_index(bound);
1571 ir_node *pred = skip_Proj(idx);
1574 if (idx == get_Bound_lower(bound))
1576 else if (is_Bound(pred)) {
1578 * idx was Bounds checked in the same MacroBlock previously,
1579 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1581 ir_node *lower = get_Bound_lower(bound);
1582 ir_node *upper = get_Bound_upper(bound);
1583 if (get_Bound_lower(pred) == lower &&
1584 get_Bound_upper(pred) == upper &&
1585 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
1587 * One could expect that we simply return the previous
1588 * Bound here. However, this would be wrong, as we could
1589 * add an exception Proj to a new location then.
1590 * So, we must turn in into a tuple.
1596 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1597 switch (get_Proj_proj(proj)) {
1599 DBG_OPT_EXC_REM(proj);
1600 proj = get_Bound_mem(bound);
1602 case pn_Bound_X_except:
1603 DBG_OPT_EXC_REM(proj);
1604 proj = get_irg_bad(current_ir_graph);
1608 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1611 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1616 } /* equivalent_node_Proj_Bound */
1619 * Optimize an Exception Proj(Load) with a non-null address.
1621 static ir_node *equivalent_node_Proj_Load(ir_node *proj) {
1622 if (get_opt_ldst_only_null_ptr_exceptions()) {
1623 if (get_irn_mode(proj) == mode_X) {
1624 ir_node *load = get_Proj_pred(proj);
1626 /* get the Load address */
1627 const ir_node *addr = get_Load_ptr(load);
1628 const ir_node *confirm;
1630 if (value_not_null(addr, &confirm)) {
1631 if (get_Proj_proj(proj) == pn_Load_X_except) {
1632 DBG_OPT_EXC_REM(proj);
1633 return get_irg_bad(current_ir_graph);
1639 } /* equivalent_node_Proj_Load */
1642 * Optimize an Exception Proj(Store) with a non-null address.
1644 static ir_node *equivalent_node_Proj_Store(ir_node *proj) {
1645 if (get_opt_ldst_only_null_ptr_exceptions()) {
1646 if (get_irn_mode(proj) == mode_X) {
1647 ir_node *store = get_Proj_pred(proj);
1649 /* get the load/store address */
1650 const ir_node *addr = get_Store_ptr(store);
1651 const ir_node *confirm;
1653 if (value_not_null(addr, &confirm)) {
1654 if (get_Proj_proj(proj) == pn_Store_X_except) {
1655 DBG_OPT_EXC_REM(proj);
1656 return get_irg_bad(current_ir_graph);
1662 } /* equivalent_node_Proj_Store */
1665 * Does all optimizations on nodes that must be done on it's Proj's
1666 * because of creating new nodes.
1668 static ir_node *equivalent_node_Proj(ir_node *proj) {
1669 ir_node *n = get_Proj_pred(proj);
1671 if (get_irn_mode(proj) == mode_X) {
1672 if (is_Block_dead(get_nodes_block(n))) {
1673 /* Remove dead control flow -- early gigo(). */
1674 return get_irg_bad(current_ir_graph);
1677 if (n->op->ops.equivalent_node_Proj)
1678 return n->op->ops.equivalent_node_Proj(proj);
1680 } /* equivalent_node_Proj */
1685 static ir_node *equivalent_node_Id(ir_node *n) {
1692 DBG_OPT_ID(oldn, n);
1694 } /* equivalent_node_Id */
1699 static ir_node *equivalent_node_Mux(ir_node *n)
1701 ir_node *oldn = n, *sel = get_Mux_sel(n);
1702 tarval *ts = value_of(sel);
1704 /* Mux(true, f, t) == t */
1705 if (ts == tarval_b_true) {
1706 n = get_Mux_true(n);
1707 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1709 /* Mux(false, f, t) == f */
1710 else if (ts == tarval_b_false) {
1711 n = get_Mux_false(n);
1712 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1714 /* Mux(v, x, x) == x */
1715 else if (get_Mux_false(n) == get_Mux_true(n)) {
1716 n = get_Mux_true(n);
1717 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1719 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1720 ir_node *cmp = get_Proj_pred(sel);
1721 long proj_nr = get_Proj_proj(sel);
1722 ir_node *f = get_Mux_false(n);
1723 ir_node *t = get_Mux_true(n);
1726 * Note further that these optimization work even for floating point
1727 * with NaN's because -NaN == NaN.
1728 * However, if +0 and -0 is handled differently, we cannot use the first one.
1731 ir_node *const cmp_l = get_Cmp_left(cmp);
1732 ir_node *const cmp_r = get_Cmp_right(cmp);
1736 if ((cmp_l == t && cmp_r == f) || /* Psi(t == f, t, f) -> f */
1737 (cmp_l == f && cmp_r == t)) { /* Psi(f == t, t, f) -> f */
1739 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1746 if ((cmp_l == t && cmp_r == f) || /* Psi(t != f, t, f) -> t */
1747 (cmp_l == f && cmp_r == t)) { /* Psi(f != t, t, f) -> t */
1749 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1756 * Note: normalization puts the constant on the right side,
1757 * so we check only one case.
1759 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1760 /* Mux(t CMP 0, X, t) */
1761 if (is_Minus(f) && get_Minus_op(f) == t) {
1762 /* Mux(t CMP 0, -t, t) */
1763 if (proj_nr == pn_Cmp_Eq) {
1764 /* Mux(t == 0, -t, t) ==> -t */
1766 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1767 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1768 /* Mux(t != 0, -t, t) ==> t */
1770 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1777 } /* equivalent_node_Mux */
1780 * Returns a equivalent node of a Psi: if a condition is true
1781 * and all previous conditions are false we know its value.
1782 * If all conditions are false its value is the default one.
1784 static ir_node *equivalent_node_Psi(ir_node *n) {
1786 return equivalent_node_Mux(n);
1788 } /* equivalent_node_Psi */
1791 * Remove Confirm nodes if setting is on.
1792 * Replace Confirms(x, '=', Constlike) by Constlike.
1794 static ir_node *equivalent_node_Confirm(ir_node *n) {
1795 ir_node *pred = get_Confirm_value(n);
1796 pn_Cmp pnc = get_Confirm_cmp(n);
1798 while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1800 * rare case: two identical Confirms one after another,
1801 * replace the second one with the first.
1804 pred = get_Confirm_value(n);
1805 pnc = get_Confirm_cmp(n);
1807 if (get_opt_remove_confirm())
1808 return get_Confirm_value(n);
1813 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1814 * perform no actual computation, as, e.g., the Id nodes. It does not create
1815 * new nodes. It is therefore safe to free n if the node returned is not n.
1816 * If a node returns a Tuple we can not just skip it. If the size of the
1817 * in array fits, we transform n into a tuple (e.g., Div).
1819 ir_node *equivalent_node(ir_node *n) {
1820 if (n->op->ops.equivalent_node)
1821 return n->op->ops.equivalent_node(n);
1823 } /* equivalent_node */
1826 * Sets the default equivalent node operation for an ir_op_ops.
1828 * @param code the opcode for the default operation
1829 * @param ops the operations initialized
1834 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1838 ops->equivalent_node = equivalent_node_##a; \
1840 #define CASE_PROJ(a) \
1842 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1885 } /* firm_set_default_equivalent_node */
1888 * Returns non-zero if a node is a Phi node
1889 * with all predecessors constant.
1891 static int is_const_Phi(ir_node *n) {
1894 if (! is_Phi(n) || get_irn_arity(n) == 0)
1896 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1897 if (! is_Const(get_irn_n(n, i)))
1900 } /* is_const_Phi */
1902 typedef tarval *(*tarval_sub_type)(tarval *a, tarval *b, ir_mode *mode);
1903 typedef tarval *(*tarval_binop_type)(tarval *a, tarval *b);
1906 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1908 static tarval *do_eval(tarval *(*eval)(), tarval *a, tarval *b, ir_mode *mode) {
1909 if (eval == tarval_sub) {
1910 tarval_sub_type func = (tarval_sub_type)eval;
1912 return func(a, b, mode);
1914 tarval_binop_type func = (tarval_binop_type)eval;
1921 * Apply an evaluator on a binop with a constant operators (and one Phi).
1923 * @param phi the Phi node
1924 * @param other the other operand
1925 * @param eval an evaluator function
1926 * @param mode the mode of the result, may be different from the mode of the Phi!
1927 * @param left if non-zero, other is the left operand, else the right
1929 * @return a new Phi node if the conversion was successful, NULL else
1931 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(), ir_mode *mode, int left) {
1936 int i, n = get_irn_arity(phi);
1938 NEW_ARR_A(void *, res, n);
1940 for (i = 0; i < n; ++i) {
1941 pred = get_irn_n(phi, i);
1942 tv = get_Const_tarval(pred);
1943 tv = do_eval(eval, other, tv, mode);
1945 if (tv == tarval_bad) {
1946 /* folding failed, bad */
1952 for (i = 0; i < n; ++i) {
1953 pred = get_irn_n(phi, i);
1954 tv = get_Const_tarval(pred);
1955 tv = do_eval(eval, tv, other, mode);
1957 if (tv == tarval_bad) {
1958 /* folding failed, bad */
1964 irg = current_ir_graph;
1965 for (i = 0; i < n; ++i) {
1966 pred = get_irn_n(phi, i);
1967 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1968 mode, res[i], get_Const_type(pred));
1970 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1971 } /* apply_binop_on_phi */
1974 * Apply an evaluator on a binop with two constant Phi.
1976 * @param a the left Phi node
1977 * @param b the right Phi node
1978 * @param eval an evaluator function
1979 * @param mode the mode of the result, may be different from the mode of the Phi!
1981 * @return a new Phi node if the conversion was successful, NULL else
1983 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(), ir_mode *mode) {
1984 tarval *tv_l, *tv_r, *tv;
1990 if (get_nodes_block(a) != get_nodes_block(b))
1993 n = get_irn_arity(a);
1994 NEW_ARR_A(void *, res, n);
1996 for (i = 0; i < n; ++i) {
1997 pred = get_irn_n(a, i);
1998 tv_l = get_Const_tarval(pred);
1999 pred = get_irn_n(b, i);
2000 tv_r = get_Const_tarval(pred);
2001 tv = do_eval(eval, tv_l, tv_r, mode);
2003 if (tv == tarval_bad) {
2004 /* folding failed, bad */
2009 irg = current_ir_graph;
2010 for (i = 0; i < n; ++i) {
2011 pred = get_irn_n(a, i);
2012 res[i] = new_r_Const_type(irg, get_irg_start_block(irg), mode, res[i], get_Const_type(pred));
2014 return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode);
2015 } /* apply_binop_on_2_phis */
2018 * Apply an evaluator on a unop with a constant operator (a Phi).
2020 * @param phi the Phi node
2021 * @param eval an evaluator function
2023 * @return a new Phi node if the conversion was successful, NULL else
2025 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
2031 int i, n = get_irn_arity(phi);
2033 NEW_ARR_A(void *, res, n);
2034 for (i = 0; i < n; ++i) {
2035 pred = get_irn_n(phi, i);
2036 tv = get_Const_tarval(pred);
2039 if (tv == tarval_bad) {
2040 /* folding failed, bad */
2045 mode = get_irn_mode(phi);
2046 irg = current_ir_graph;
2047 for (i = 0; i < n; ++i) {
2048 pred = get_irn_n(phi, i);
2049 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
2050 mode, res[i], get_Const_type(pred));
2052 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2053 } /* apply_unop_on_phi */
2056 * Apply a conversion on a constant operator (a Phi).
2058 * @param phi the Phi node
2060 * @return a new Phi node if the conversion was successful, NULL else
2062 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
2067 int i, n = get_irn_arity(phi);
2069 NEW_ARR_A(void *, res, n);
2070 for (i = 0; i < n; ++i) {
2071 pred = get_irn_n(phi, i);
2072 tv = get_Const_tarval(pred);
2073 tv = tarval_convert_to(tv, mode);
2075 if (tv == tarval_bad) {
2076 /* folding failed, bad */
2081 irg = current_ir_graph;
2082 for (i = 0; i < n; ++i) {
2083 pred = get_irn_n(phi, i);
2084 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
2085 mode, res[i], get_Const_type(pred));
2087 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2088 } /* apply_conv_on_phi */
2091 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
2092 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
2093 * If possible, remove the Conv's.
2095 static ir_node *transform_node_AddSub(ir_node *n) {
2096 ir_mode *mode = get_irn_mode(n);
2098 if (mode_is_reference(mode)) {
2099 ir_node *left = get_binop_left(n);
2100 ir_node *right = get_binop_right(n);
2101 unsigned ref_bits = get_mode_size_bits(mode);
2103 if (is_Conv(left)) {
2104 ir_mode *lmode = get_irn_mode(left);
2105 unsigned bits = get_mode_size_bits(lmode);
2107 if (ref_bits == bits &&
2108 mode_is_int(lmode) &&
2109 get_mode_arithmetic(lmode) == irma_twos_complement) {
2110 ir_node *pre = get_Conv_op(left);
2111 ir_mode *pre_mode = get_irn_mode(pre);
2113 if (mode_is_int(pre_mode) &&
2114 get_mode_size_bits(pre_mode) == bits &&
2115 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2116 /* ok, this conv just changes to sign, moreover the calculation
2117 * is done with same number of bits as our address mode, so
2118 * we can ignore the conv as address calculation can be viewed
2119 * as either signed or unsigned
2121 set_binop_left(n, pre);
2126 if (is_Conv(right)) {
2127 ir_mode *rmode = get_irn_mode(right);
2128 unsigned bits = get_mode_size_bits(rmode);
2130 if (ref_bits == bits &&
2131 mode_is_int(rmode) &&
2132 get_mode_arithmetic(rmode) == irma_twos_complement) {
2133 ir_node *pre = get_Conv_op(right);
2134 ir_mode *pre_mode = get_irn_mode(pre);
2136 if (mode_is_int(pre_mode) &&
2137 get_mode_size_bits(pre_mode) == bits &&
2138 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2139 /* ok, this conv just changes to sign, moreover the calculation
2140 * is done with same number of bits as our address mode, so
2141 * we can ignore the conv as address calculation can be viewed
2142 * as either signed or unsigned
2144 set_binop_right(n, pre);
2149 /* let address arithmetic use unsigned modes */
2150 if (is_Const(right)) {
2151 ir_mode *rmode = get_irn_mode(right);
2153 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2154 /* convert a AddP(P, *s) into AddP(P, *u) */
2155 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2157 ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
2158 set_binop_right(n, pre);
2164 } /* transform_node_AddSub */
2166 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2168 if (is_Const(b) && is_const_Phi(a)) { \
2169 /* check for Op(Phi, Const) */ \
2170 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2172 else if (is_Const(a) && is_const_Phi(b)) { \
2173 /* check for Op(Const, Phi) */ \
2174 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2176 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2177 /* check for Op(Phi, Phi) */ \
2178 c = apply_binop_on_2_phis(a, b, eval, mode); \
2181 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2185 #define HANDLE_UNOP_PHI(eval, a, c) \
2187 if (is_const_Phi(a)) { \
2188 /* check for Op(Phi) */ \
2189 c = apply_unop_on_phi(a, eval); \
2191 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2197 * Do the AddSub optimization, then Transform
2198 * Constant folding on Phi
2199 * Add(a,a) -> Mul(a, 2)
2200 * Add(Mul(a, x), a) -> Mul(a, x+1)
2201 * if the mode is integer or float.
2202 * Transform Add(a,-b) into Sub(a,b).
2203 * Reassociation might fold this further.
2205 static ir_node *transform_node_Add(ir_node *n) {
2207 ir_node *a, *b, *c, *oldn = n;
2209 n = transform_node_AddSub(n);
2211 a = get_Add_left(n);
2212 b = get_Add_right(n);
2214 mode = get_irn_mode(n);
2216 if (mode_is_reference(mode)) {
2217 ir_mode *lmode = get_irn_mode(a);
2219 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2220 /* an Add(a, NULL) is a hidden Conv */
2221 dbg_info *dbg = get_irn_dbg_info(n);
2222 return new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2226 HANDLE_BINOP_PHI(tarval_add, a, b, c, mode);
2228 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2229 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2232 if (mode_is_num(mode)) {
2233 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2234 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
2235 ir_node *block = get_nodes_block(n);
2238 get_irn_dbg_info(n),
2242 new_r_Const_long(current_ir_graph, block, mode, 2),
2244 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2249 get_irn_dbg_info(n),
2255 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2260 get_irn_dbg_info(n),
2266 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2269 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2270 /* Here we rely on constants be on the RIGHT side */
2272 ir_node *op = get_Not_op(a);
2274 if (is_Const(b) && is_Const_one(b)) {
2276 ir_node *blk = get_irn_n(n, -1);
2277 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2278 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2283 ir_node *blk = get_irn_n(n, -1);
2284 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2285 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2290 ir_node *op = get_Not_op(b);
2294 ir_node *blk = get_irn_n(n, -1);
2295 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2296 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2303 } /* transform_node_Add */
2306 * returns -cnst or NULL if impossible
2308 static ir_node *const_negate(ir_node *cnst) {
2309 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2310 dbg_info *dbgi = get_irn_dbg_info(cnst);
2311 ir_graph *irg = get_irn_irg(cnst);
2312 ir_node *block = get_nodes_block(cnst);
2313 ir_mode *mode = get_irn_mode(cnst);
2314 if (tv == tarval_bad) return NULL;
2315 return new_rd_Const(dbgi, irg, block, mode, tv);
2319 * Do the AddSub optimization, then Transform
2320 * Constant folding on Phi
2321 * Sub(0,a) -> Minus(a)
2322 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2323 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2324 * Sub(Add(a, x), x) -> a
2325 * Sub(x, Add(x, a)) -> -a
2326 * Sub(x, Const) -> Add(x, -Const)
2328 static ir_node *transform_node_Sub(ir_node *n) {
2333 n = transform_node_AddSub(n);
2335 a = get_Sub_left(n);
2336 b = get_Sub_right(n);
2338 mode = get_irn_mode(n);
2340 if (mode_is_int(mode)) {
2341 ir_mode *lmode = get_irn_mode(a);
2343 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2344 /* a Sub(a, NULL) is a hidden Conv */
2345 dbg_info *dbg = get_irn_dbg_info(n);
2346 return new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2351 HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode);
2353 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2354 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2357 if (is_Const(b) && get_irn_mode(b) != mode_P) {
2358 /* a - C -> a + (-C) */
2359 ir_node *cnst = const_negate(b);
2361 ir_node *block = get_nodes_block(n);
2362 dbg_info *dbgi = get_irn_dbg_info(n);
2363 ir_graph *irg = get_irn_irg(n);
2365 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2366 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2371 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2372 ir_graph *irg = current_ir_graph;
2373 dbg_info *dbg = get_irn_dbg_info(n);
2374 ir_node *block = get_nodes_block(n);
2375 ir_node *left = get_Minus_op(a);
2376 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2378 n = new_rd_Minus(dbg, irg, block, add, mode);
2379 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2381 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2382 ir_graph *irg = current_ir_graph;
2383 dbg_info *dbg = get_irn_dbg_info(n);
2384 ir_node *block = get_nodes_block(n);
2385 ir_node *right = get_Minus_op(b);
2387 n = new_rd_Add(dbg, irg, block, a, right, mode);
2388 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2390 } else if (is_Sub(b)) { /* a - (b - c) -> a + (c - b) */
2391 ir_graph *irg = current_ir_graph;
2392 dbg_info *s_dbg = get_irn_dbg_info(b);
2393 ir_node *s_block = get_nodes_block(b);
2394 ir_node *s_left = get_Sub_right(b);
2395 ir_node *s_right = get_Sub_left(b);
2396 ir_mode *s_mode = get_irn_mode(b);
2397 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_left, s_right, s_mode);
2398 dbg_info *a_dbg = get_irn_dbg_info(n);
2399 ir_node *a_block = get_nodes_block(n);
2401 n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode);
2402 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2404 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2405 ir_node *m_right = get_Mul_right(b);
2406 if (is_Const(m_right)) {
2407 ir_node *cnst2 = const_negate(m_right);
2408 if (cnst2 != NULL) {
2409 ir_graph *irg = current_ir_graph;
2410 dbg_info *m_dbg = get_irn_dbg_info(b);
2411 ir_node *m_block = get_nodes_block(b);
2412 ir_node *m_left = get_Mul_left(b);
2413 ir_mode *m_mode = get_irn_mode(b);
2414 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2415 dbg_info *a_dbg = get_irn_dbg_info(n);
2416 ir_node *a_block = get_nodes_block(n);
2418 n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode);
2419 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2425 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2426 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2428 get_irn_dbg_info(n),
2433 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2437 if (mode_wrap_around(mode)) {
2438 ir_node *left = get_Add_left(a);
2439 ir_node *right = get_Add_right(a);
2441 /* FIXME: Does the Conv's work only for two complement or generally? */
2443 if (mode != get_irn_mode(right)) {
2444 /* This Sub is an effective Cast */
2445 right = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), right, mode);
2448 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2450 } else if (right == b) {
2451 if (mode != get_irn_mode(left)) {
2452 /* This Sub is an effective Cast */
2453 left = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), left, mode);
2456 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2462 if (mode_wrap_around(mode)) {
2463 ir_node *left = get_Add_left(b);
2464 ir_node *right = get_Add_right(b);
2466 /* FIXME: Does the Conv's work only for two complement or generally? */
2468 ir_mode *r_mode = get_irn_mode(right);
2470 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), right, r_mode);
2471 if (mode != r_mode) {
2472 /* This Sub is an effective Cast */
2473 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2475 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2477 } else if (right == a) {
2478 ir_mode *l_mode = get_irn_mode(left);
2480 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), left, l_mode);
2481 if (mode != l_mode) {
2482 /* This Sub is an effective Cast */
2483 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2485 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2490 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2491 ir_mode *mode = get_irn_mode(a);
2493 if (mode == get_irn_mode(b)) {
2495 ir_node *op_a = get_Conv_op(a);
2496 ir_node *op_b = get_Conv_op(b);
2498 /* check if it's allowed to skip the conv */
2499 ma = get_irn_mode(op_a);
2500 mb = get_irn_mode(op_b);
2502 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2503 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2506 set_Sub_right(n, b);
2512 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2513 if (!is_reassoc_running() && is_Mul(a)) {
2514 ir_node *ma = get_Mul_left(a);
2515 ir_node *mb = get_Mul_right(a);
2518 ir_node *blk = get_irn_n(n, -1);
2520 get_irn_dbg_info(n),
2521 current_ir_graph, blk,
2524 get_irn_dbg_info(n),
2525 current_ir_graph, blk,
2527 new_r_Const_long(current_ir_graph, blk, mode, 1),
2530 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2532 } else if (mb == b) {
2533 ir_node *blk = get_irn_n(n, -1);
2535 get_irn_dbg_info(n),
2536 current_ir_graph, blk,
2539 get_irn_dbg_info(n),
2540 current_ir_graph, blk,
2542 new_r_Const_long(current_ir_graph, blk, mode, 1),
2545 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2549 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2550 ir_node *x = get_Sub_left(a);
2551 ir_node *y = get_Sub_right(a);
2552 ir_node *blk = get_irn_n(n, -1);
2553 ir_mode *m_b = get_irn_mode(b);
2554 ir_mode *m_y = get_irn_mode(y);
2558 /* Determine the right mode for the Add. */
2561 else if (mode_is_reference(m_b))
2563 else if (mode_is_reference(m_y))
2567 * Both modes are different but none is reference,
2568 * happens for instance in SubP(SubP(P, Iu), Is).
2569 * We have two possibilities here: Cast or ignore.
2570 * Currently we ignore this case.
2575 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2577 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2578 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2582 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2583 if (is_Const(a) && is_Not(b)) {
2584 /* c - ~X = X + (c+1) */
2585 tarval *tv = get_Const_tarval(a);
2587 tv = tarval_add(tv, get_mode_one(mode));
2588 if (tv != tarval_bad) {
2589 ir_node *blk = get_irn_n(n, -1);
2590 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2591 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2592 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2598 } /* transform_node_Sub */
2601 * Several transformation done on n*n=2n bits mul.
2602 * These transformations must be done here because new nodes may be produced.
2604 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2606 ir_node *a = get_Mul_left(n);
2607 ir_node *b = get_Mul_right(n);
2608 tarval *ta = value_of(a);
2609 tarval *tb = value_of(b);
2610 ir_mode *smode = get_irn_mode(a);
2612 if (ta == get_mode_one(smode)) {
2613 /* (L)1 * (L)b = (L)b */
2614 ir_node *blk = get_irn_n(n, -1);
2615 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2616 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2619 else if (ta == get_mode_minus_one(smode)) {
2620 /* (L)-1 * (L)b = (L)b */
2621 ir_node *blk = get_irn_n(n, -1);
2622 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2623 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2624 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2627 if (tb == get_mode_one(smode)) {
2628 /* (L)a * (L)1 = (L)a */
2629 ir_node *blk = get_irn_n(a, -1);
2630 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2631 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2634 else if (tb == get_mode_minus_one(smode)) {
2635 /* (L)a * (L)-1 = (L)-a */
2636 ir_node *blk = get_irn_n(n, -1);
2637 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2638 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2639 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2646 * Transform Mul(a,-1) into -a.
2647 * Do constant evaluation of Phi nodes.
2648 * Do architecture dependent optimizations on Mul nodes
2650 static ir_node *transform_node_Mul(ir_node *n) {
2651 ir_node *c, *oldn = n;
2652 ir_mode *mode = get_irn_mode(n);
2653 ir_node *a = get_Mul_left(n);
2654 ir_node *b = get_Mul_right(n);
2656 if (is_Bad(a) || is_Bad(b))
2659 if (mode != get_irn_mode(a))
2660 return transform_node_Mul2n(n, mode);
2662 HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode);
2664 if (mode_is_signed(mode)) {
2667 if (value_of(a) == get_mode_minus_one(mode))
2669 else if (value_of(b) == get_mode_minus_one(mode))
2672 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2673 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2678 if (is_Const(b)) { /* (-a) * const -> a * -const */
2679 ir_node *cnst = const_negate(b);
2681 dbg_info *dbgi = get_irn_dbg_info(n);
2682 ir_node *block = get_nodes_block(n);
2683 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2684 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2687 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2688 dbg_info *dbgi = get_irn_dbg_info(n);
2689 ir_node *block = get_nodes_block(n);
2690 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2691 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2693 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2694 ir_node *sub_l = get_Sub_left(b);
2695 ir_node *sub_r = get_Sub_right(b);
2696 dbg_info *dbgi = get_irn_dbg_info(n);
2697 ir_graph *irg = current_ir_graph;
2698 ir_node *block = get_nodes_block(n);
2699 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2700 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2701 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2704 } else if (is_Minus(b)) {
2705 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2706 ir_node *sub_l = get_Sub_left(a);
2707 ir_node *sub_r = get_Sub_right(a);
2708 dbg_info *dbgi = get_irn_dbg_info(n);
2709 ir_graph *irg = current_ir_graph;
2710 ir_node *block = get_nodes_block(n);
2711 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2712 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2713 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2717 if (get_mode_arithmetic(mode) == irma_ieee754) {
2719 tarval *tv = get_Const_tarval(a);
2720 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2721 /* 2.0 * b = b + b */
2722 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode);
2723 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2727 else if (is_Const(b)) {
2728 tarval *tv = get_Const_tarval(b);
2729 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2730 /* a * 2.0 = a + a */
2731 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode);
2732 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2737 return arch_dep_replace_mul_with_shifts(n);
2738 } /* transform_node_Mul */
2741 * Transform a Div Node.
2743 static ir_node *transform_node_Div(ir_node *n) {
2744 ir_mode *mode = get_Div_resmode(n);
2745 ir_node *a = get_Div_left(n);
2746 ir_node *b = get_Div_right(n);
2750 if (is_Const(b) && is_const_Phi(a)) {
2751 /* check for Div(Phi, Const) */
2752 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2754 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2758 else if (is_Const(a) && is_const_Phi(b)) {
2759 /* check for Div(Const, Phi) */
2760 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2762 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2766 else if (is_const_Phi(a) && is_const_Phi(b)) {
2767 /* check for Div(Phi, Phi) */
2768 value = apply_binop_on_2_phis(a, b, tarval_div, mode);
2770 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2777 if (tv != tarval_bad) {
2778 value = new_Const(get_tarval_mode(tv), tv);
2780 DBG_OPT_CSTEVAL(n, value);
2783 ir_node *a = get_Div_left(n);
2784 ir_node *b = get_Div_right(n);
2785 const ir_node *dummy;
2787 if (a == b && value_not_zero(a, &dummy)) {
2788 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2789 value = new_Const(mode, get_mode_one(mode));
2790 DBG_OPT_CSTEVAL(n, value);
2793 if (mode_is_signed(mode) && is_Const(b)) {
2794 tarval *tv = get_Const_tarval(b);
2796 if (tv == get_mode_minus_one(mode)) {
2798 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2799 DBG_OPT_CSTEVAL(n, value);
2803 /* Try architecture dependent optimization */
2804 value = arch_dep_replace_div_by_const(n);
2812 /* Turn Div into a tuple (mem, jmp, bad, value) */
2813 mem = get_Div_mem(n);
2814 blk = get_irn_n(n, -1);
2816 /* skip a potential Pin */
2818 mem = get_Pin_op(mem);
2819 turn_into_tuple(n, pn_Div_max);
2820 set_Tuple_pred(n, pn_Div_M, mem);
2821 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2822 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2823 set_Tuple_pred(n, pn_Div_res, value);
2826 } /* transform_node_Div */
2829 * Transform a Mod node.
2831 static ir_node *transform_node_Mod(ir_node *n) {
2832 ir_mode *mode = get_Mod_resmode(n);
2833 ir_node *a = get_Mod_left(n);
2834 ir_node *b = get_Mod_right(n);
2838 if (is_Const(b) && is_const_Phi(a)) {
2839 /* check for Div(Phi, Const) */
2840 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2842 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2846 else if (is_Const(a) && is_const_Phi(b)) {
2847 /* check for Div(Const, Phi) */
2848 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2850 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2854 else if (is_const_Phi(a) && is_const_Phi(b)) {
2855 /* check for Div(Phi, Phi) */
2856 value = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2858 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2865 if (tv != tarval_bad) {
2866 value = new_Const(get_tarval_mode(tv), tv);
2868 DBG_OPT_CSTEVAL(n, value);
2871 ir_node *a = get_Mod_left(n);
2872 ir_node *b = get_Mod_right(n);
2873 const ir_node *dummy;
2875 if (a == b && value_not_zero(a, &dummy)) {
2876 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2877 value = new_Const(mode, get_mode_null(mode));
2878 DBG_OPT_CSTEVAL(n, value);
2881 if (mode_is_signed(mode) && is_Const(b)) {
2882 tarval *tv = get_Const_tarval(b);
2884 if (tv == get_mode_minus_one(mode)) {
2886 value = new_Const(mode, get_mode_null(mode));
2887 DBG_OPT_CSTEVAL(n, value);
2891 /* Try architecture dependent optimization */
2892 value = arch_dep_replace_mod_by_const(n);
2900 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2901 mem = get_Mod_mem(n);
2902 blk = get_irn_n(n, -1);
2904 /* skip a potential Pin */
2906 mem = get_Pin_op(mem);
2907 turn_into_tuple(n, pn_Mod_max);
2908 set_Tuple_pred(n, pn_Mod_M, mem);
2909 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2910 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2911 set_Tuple_pred(n, pn_Mod_res, value);
2914 } /* transform_node_Mod */
2917 * Transform a DivMod node.
2919 static ir_node *transform_node_DivMod(ir_node *n) {
2920 const ir_node *dummy;
2921 ir_node *a = get_DivMod_left(n);
2922 ir_node *b = get_DivMod_right(n);
2923 ir_mode *mode = get_DivMod_resmode(n);
2928 if (is_Const(b) && is_const_Phi(a)) {
2929 /* check for Div(Phi, Const) */
2930 va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2931 vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2933 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2934 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2938 else if (is_Const(a) && is_const_Phi(b)) {
2939 /* check for Div(Const, Phi) */
2940 va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2941 vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2943 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2944 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2948 else if (is_const_Phi(a) && is_const_Phi(b)) {
2949 /* check for Div(Phi, Phi) */
2950 va = apply_binop_on_2_phis(a, b, tarval_div, mode);
2951 vb = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2953 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2954 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2961 if (tb != tarval_bad) {
2962 if (tb == get_mode_one(get_tarval_mode(tb))) {
2964 vb = new_Const(mode, get_mode_null(mode));
2965 DBG_OPT_CSTEVAL(n, vb);
2967 } else if (ta != tarval_bad) {
2968 tarval *resa, *resb;
2969 resa = tarval_div(ta, tb);
2970 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2971 Jmp for X result!? */
2972 resb = tarval_mod(ta, tb);
2973 if (resb == tarval_bad) return n; /* Causes exception! */
2974 va = new_Const(mode, resa);
2975 vb = new_Const(mode, resb);
2976 DBG_OPT_CSTEVAL(n, va);
2977 DBG_OPT_CSTEVAL(n, vb);
2979 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
2980 va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2981 vb = new_Const(mode, get_mode_null(mode));
2982 DBG_OPT_CSTEVAL(n, va);
2983 DBG_OPT_CSTEVAL(n, vb);
2985 } else { /* Try architecture dependent optimization */
2988 arch_dep_replace_divmod_by_const(&va, &vb, n);
2989 evaluated = va != NULL;
2991 } else if (a == b) {
2992 if (value_not_zero(a, &dummy)) {
2994 va = new_Const(mode, get_mode_one(mode));
2995 vb = new_Const(mode, get_mode_null(mode));
2996 DBG_OPT_CSTEVAL(n, va);
2997 DBG_OPT_CSTEVAL(n, vb);
3000 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
3003 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
3004 /* 0 / non-Const = 0 */
3009 if (evaluated) { /* replace by tuple */
3013 mem = get_DivMod_mem(n);
3014 /* skip a potential Pin */
3016 mem = get_Pin_op(mem);
3018 blk = get_irn_n(n, -1);
3019 turn_into_tuple(n, pn_DivMod_max);
3020 set_Tuple_pred(n, pn_DivMod_M, mem);
3021 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
3022 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
3023 set_Tuple_pred(n, pn_DivMod_res_div, va);
3024 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
3028 } /* transform_node_DivMod */
3031 * Optimize x / c to x * (1/c)
3033 static ir_node *transform_node_Quot(ir_node *n) {
3034 ir_mode *mode = get_Quot_resmode(n);
3037 if (get_mode_arithmetic(mode) == irma_ieee754) {
3038 ir_node *b = get_Quot_right(n);
3041 tarval *tv = get_Const_tarval(b);
3045 * Floating point constant folding might be disabled here to
3047 * However, as we check for exact result, doing it is safe.
3050 rem = tarval_enable_fp_ops(1);
3051 tv = tarval_quo(get_mode_one(mode), tv);
3052 (void)tarval_enable_fp_ops(rem);
3054 /* Do the transformation if the result is either exact or we are not
3055 using strict rules. */
3056 if (tv != tarval_bad &&
3057 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
3058 ir_node *blk = get_irn_n(n, -1);
3059 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3060 ir_node *a = get_Quot_left(n);
3061 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
3062 ir_node *mem = get_Quot_mem(n);
3064 /* skip a potential Pin */
3066 mem = get_Pin_op(mem);
3067 turn_into_tuple(n, pn_Quot_max);
3068 set_Tuple_pred(n, pn_Quot_M, mem);
3069 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
3070 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
3071 set_Tuple_pred(n, pn_Quot_res, m);
3072 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
3077 } /* transform_node_Quot */
3080 * Optimize Abs(x) into x if x is Confirmed >= 0
3081 * Optimize Abs(x) into -x if x is Confirmed <= 0
3082 * Optimize Abs(-x) int Abs(x)
3084 static ir_node *transform_node_Abs(ir_node *n) {
3085 ir_node *c, *oldn = n;
3086 ir_node *a = get_Abs_op(n);
3089 HANDLE_UNOP_PHI(tarval_abs, a, c);
3091 switch (classify_value_sign(a)) {
3092 case value_classified_negative:
3093 mode = get_irn_mode(n);
3096 * We can replace the Abs by -x here.
3097 * We even could add a new Confirm here
3098 * (if not twos complement)
3100 * Note that -x would create a new node, so we could
3101 * not run it in the equivalent_node() context.
3103 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
3104 get_nodes_block(n), a, mode);
3106 DBG_OPT_CONFIRM(oldn, n);
3108 case value_classified_positive:
3109 /* n is positive, Abs is not needed */
3112 DBG_OPT_CONFIRM(oldn, n);
3118 /* Abs(-x) = Abs(x) */
3119 mode = get_irn_mode(n);
3120 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph,
3121 get_nodes_block(n), get_Minus_op(a), mode);
3122 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
3126 } /* transform_node_Abs */
3129 * Optimize -a CMP -b into b CMP a.
3130 * This works only for for modes where unary Minus
3132 * Note that two-complement integers can Overflow
3133 * so it will NOT work.
3135 * For == and != can be handled in Proj(Cmp)
3137 static ir_node *transform_node_Cmp(ir_node *n) {
3139 ir_node *left = get_Cmp_left(n);
3140 ir_node *right = get_Cmp_right(n);
3142 if (is_Minus(left) && is_Minus(right) &&
3143 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
3144 left = get_Minus_op(left);
3145 right = get_Minus_op(right);
3146 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph,
3147 get_nodes_block(n), left, right);
3148 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP);
3151 } /* transform_node_Cmp */
3155 * Transform a Cond node.
3157 * Replace the Cond by a Jmp if it branches on a constant
3160 static ir_node *transform_node_Cond(ir_node *n) {
3163 ir_node *a = get_Cond_selector(n);
3164 tarval *ta = value_of(a);
3166 /* we need block info which is not available in floating irgs */
3167 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
3170 if ((ta != tarval_bad) &&
3171 (get_irn_mode(a) == mode_b) &&
3172 (get_opt_unreachable_code())) {
3173 /* It's a boolean Cond, branching on a boolean constant.
3174 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3175 ir_node *blk = get_nodes_block(n);
3176 jmp = new_r_Jmp(current_ir_graph, blk);
3177 turn_into_tuple(n, pn_Cond_max);
3178 if (ta == tarval_b_true) {
3179 set_Tuple_pred(n, pn_Cond_false, new_Bad());
3180 set_Tuple_pred(n, pn_Cond_true, jmp);
3182 set_Tuple_pred(n, pn_Cond_false, jmp);
3183 set_Tuple_pred(n, pn_Cond_true, new_Bad());
3185 /* We might generate an endless loop, so keep it alive. */
3186 add_End_keepalive(get_irg_end(current_ir_graph), blk);
3189 } /* transform_node_Cond */
3192 * Prototype of a recursive transform function
3193 * for bitwise distributive transformations.
3195 typedef ir_node* (*recursive_transform)(ir_node *n);
3198 * makes use of distributive laws for and, or, eor
3199 * and(a OP c, b OP c) -> and(a, b) OP c
3200 * note, might return a different op than n
3202 static ir_node *transform_bitwise_distributive(ir_node *n,
3203 recursive_transform trans_func)
3206 ir_node *a = get_binop_left(n);
3207 ir_node *b = get_binop_right(n);
3208 ir_op *op = get_irn_op(a);
3209 ir_op *op_root = get_irn_op(n);
3211 if(op != get_irn_op(b))
3214 if (op == op_Conv) {
3215 ir_node *a_op = get_Conv_op(a);
3216 ir_node *b_op = get_Conv_op(b);
3217 ir_mode *a_mode = get_irn_mode(a_op);
3218 ir_mode *b_mode = get_irn_mode(b_op);
3219 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3220 ir_node *blk = get_irn_n(n, -1);
3223 set_binop_left(n, a_op);
3224 set_binop_right(n, b_op);
3225 set_irn_mode(n, a_mode);
3227 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
3229 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3235 /* nothing to gain here */
3239 if (op == op_Shrs || op == op_Shr || op == op_Shl
3240 || op == op_And || op == op_Or || op == op_Eor) {
3241 ir_node *a_left = get_binop_left(a);
3242 ir_node *a_right = get_binop_right(a);
3243 ir_node *b_left = get_binop_left(b);
3244 ir_node *b_right = get_binop_right(b);
3246 ir_node *op1 = NULL;
3247 ir_node *op2 = NULL;
3249 if (is_op_commutative(op)) {
3250 if (a_left == b_left) {
3254 } else if(a_left == b_right) {
3258 } else if(a_right == b_left) {
3264 if(a_right == b_right) {
3271 /* (a sop c) & (b sop c) => (a & b) sop c */
3272 ir_node *blk = get_irn_n(n, -1);
3274 ir_node *new_n = exact_copy(n);
3275 set_binop_left(new_n, op1);
3276 set_binop_right(new_n, op2);
3277 new_n = trans_func(new_n);
3279 if(op_root == op_Eor && op == op_Or) {
3280 dbg_info *dbgi = get_irn_dbg_info(n);
3281 ir_graph *irg = current_ir_graph;
3282 ir_mode *mode = get_irn_mode(c);
3284 c = new_rd_Not(dbgi, irg, blk, c, mode);
3285 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
3288 set_nodes_block(n, blk);
3289 set_binop_left(n, new_n);
3290 set_binop_right(n, c);
3291 add_identities(current_ir_graph->value_table, n);
3294 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3305 static ir_node *transform_node_And(ir_node *n) {
3306 ir_node *c, *oldn = n;
3307 ir_node *a = get_And_left(n);
3308 ir_node *b = get_And_right(n);
3311 mode = get_irn_mode(n);
3312 HANDLE_BINOP_PHI(tarval_and, a, b, c, mode);
3314 /* we can evaluate 2 Projs of the same Cmp */
3315 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3316 ir_node *pred_a = get_Proj_pred(a);
3317 ir_node *pred_b = get_Proj_pred(b);
3318 if (pred_a == pred_b) {
3319 dbg_info *dbgi = get_irn_dbg_info(n);
3320 ir_node *block = get_nodes_block(pred_a);
3321 pn_Cmp pn_a = get_Proj_proj(a);
3322 pn_Cmp pn_b = get_Proj_proj(b);
3323 /* yes, we can simply calculate with pncs */
3324 pn_Cmp new_pnc = pn_a & pn_b;
3326 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
3331 ir_node *op = get_Not_op(b);
3333 ir_node *ba = get_And_left(op);
3334 ir_node *bb = get_And_right(op);
3336 /* it's enough to test the following cases due to normalization! */
3337 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3338 /* (a|b) & ~(a&b) = a^b */
3339 ir_node *block = get_nodes_block(n);
3341 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
3342 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3350 ir_node *op = get_Not_op(a);
3352 ir_node *aa = get_And_left(op);
3353 ir_node *ab = get_And_right(op);
3355 /* it's enough to test the following cases due to normalization! */
3356 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3357 /* (a|b) & ~(a&b) = a^b */
3358 ir_node *block = get_nodes_block(n);
3360 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
3361 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3368 ir_node *al = get_Eor_left(a);
3369 ir_node *ar = get_Eor_right(a);
3372 /* (b ^ a) & b -> ~a & b */
3373 dbg_info *dbg = get_irn_dbg_info(n);
3374 ir_node *block = get_nodes_block(n);
3376 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
3377 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
3378 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3382 /* (a ^ b) & b -> ~a & b */
3383 dbg_info *dbg = get_irn_dbg_info(n);
3384 ir_node *block = get_nodes_block(n);
3386 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
3387 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
3388 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3393 ir_node *bl = get_Eor_left(b);
3394 ir_node *br = get_Eor_right(b);
3397 /* a & (a ^ b) -> a & ~b */
3398 dbg_info *dbg = get_irn_dbg_info(n);
3399 ir_node *block = get_nodes_block(n);
3401 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
3402 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
3403 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3407 /* a & (b ^ a) -> a & ~b */
3408 dbg_info *dbg = get_irn_dbg_info(n);
3409 ir_node *block = get_nodes_block(n);
3411 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3412 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3413 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3417 if (is_Not(a) && is_Not(b)) {
3418 /* ~a & ~b = ~(a|b) */
3419 ir_node *block = get_nodes_block(n);
3420 ir_mode *mode = get_irn_mode(n);
3424 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3425 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3426 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3430 n = transform_bitwise_distributive(n, transform_node_And);
3433 } /* transform_node_And */
3438 static ir_node *transform_node_Eor(ir_node *n) {
3439 ir_node *c, *oldn = n;
3440 ir_node *a = get_Eor_left(n);
3441 ir_node *b = get_Eor_right(n);
3442 ir_mode *mode = get_irn_mode(n);
3444 HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode);
3446 /* we can evaluate 2 Projs of the same Cmp */
3447 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3448 ir_node *pred_a = get_Proj_pred(a);
3449 ir_node *pred_b = get_Proj_pred(b);
3450 if(pred_a == pred_b) {
3451 dbg_info *dbgi = get_irn_dbg_info(n);
3452 ir_node *block = get_nodes_block(pred_a);
3453 pn_Cmp pn_a = get_Proj_proj(a);
3454 pn_Cmp pn_b = get_Proj_proj(b);
3455 /* yes, we can simply calculate with pncs */
3456 pn_Cmp new_pnc = pn_a ^ pn_b;
3458 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3465 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
3466 mode, get_mode_null(mode));
3467 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3468 } else if (mode == mode_b &&
3470 is_Const(b) && is_Const_one(b) &&
3471 is_Cmp(get_Proj_pred(a))) {
3472 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3473 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3474 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3476 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3477 } else if (is_Const(b)) {
3478 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3479 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b)));
3480 ir_node *not_op = get_Not_op(a);
3481 dbg_info *dbg = get_irn_dbg_info(n);
3482 ir_graph *irg = current_ir_graph;
3483 ir_node *block = get_nodes_block(n);
3484 ir_mode *mode = get_irn_mode(n);
3485 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3487 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3488 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3489 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3492 n = transform_bitwise_distributive(n, transform_node_Eor);
3496 } /* transform_node_Eor */
3501 static ir_node *transform_node_Not(ir_node *n) {
3502 ir_node *c, *oldn = n;
3503 ir_node *a = get_Not_op(n);
3504 ir_mode *mode = get_irn_mode(n);
3506 HANDLE_UNOP_PHI(tarval_not,a,c);
3508 /* check for a boolean Not */
3509 if (mode == mode_b &&
3511 is_Cmp(get_Proj_pred(a))) {
3512 /* We negate a Cmp. The Cmp has the negated result anyways! */
3513 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3514 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3515 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3519 ir_node *eor_b = get_Eor_right(a);
3520 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3521 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b)));
3522 ir_node *eor_a = get_Eor_left(a);
3523 dbg_info *dbg = get_irn_dbg_info(n);
3524 ir_graph *irg = current_ir_graph;
3525 ir_node *block = get_nodes_block(n);
3526 ir_mode *mode = get_irn_mode(n);
3527 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3531 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3532 if (is_Minus(a)) { /* ~-x -> x + -1 */
3533 dbg_info *dbg = get_irn_dbg_info(n);
3534 ir_graph *irg = current_ir_graph;
3535 ir_node *block = get_nodes_block(n);
3536 ir_node *add_l = get_Minus_op(a);
3537 ir_node *add_r = new_rd_Const(dbg, irg, block, mode, get_mode_minus_one(mode));
3538 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3539 } else if (is_Add(a)) {
3540 ir_node *add_r = get_Add_right(a);
3541 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3542 /* ~(x + -1) = -x */
3543 ir_node *op = get_Add_left(a);
3544 ir_node *blk = get_irn_n(n, -1);
3545 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3546 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3551 } /* transform_node_Not */
3554 * Transform a Minus.
3558 * -(a >>u (size-1)) = a >>s (size-1)
3559 * -(a >>s (size-1)) = a >>u (size-1)
3560 * -(a * const) -> a * -const
3562 static ir_node *transform_node_Minus(ir_node *n) {
3563 ir_node *c, *oldn = n;
3564 ir_node *a = get_Minus_op(n);
3567 HANDLE_UNOP_PHI(tarval_neg,a,c);
3569 mode = get_irn_mode(a);
3570 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3571 /* the following rules are only to twos-complement */
3574 ir_node *op = get_Not_op(a);
3575 tarval *tv = get_mode_one(mode);
3576 ir_node *blk = get_irn_n(n, -1);
3577 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3578 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3579 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3583 ir_node *c = get_Shr_right(a);
3586 tarval *tv = get_Const_tarval(c);
3588 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3589 /* -(a >>u (size-1)) = a >>s (size-1) */
3590 ir_node *v = get_Shr_left(a);
3592 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3593 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3599 ir_node *c = get_Shrs_right(a);
3602 tarval *tv = get_Const_tarval(c);
3604 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3605 /* -(a >>s (size-1)) = a >>u (size-1) */
3606 ir_node *v = get_Shrs_left(a);
3608 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3609 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3616 /* - (a-b) = b - a */
3617 ir_node *la = get_Sub_left(a);
3618 ir_node *ra = get_Sub_right(a);
3619 ir_node *blk = get_irn_n(n, -1);
3621 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3622 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3626 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3627 ir_node *mul_l = get_Mul_left(a);
3628 ir_node *mul_r = get_Mul_right(a);
3629 if (is_Const(mul_r)) {
3630 tarval *tv = tarval_neg(get_Const_tarval(mul_r));
3631 if(tv != tarval_bad) {
3632 ir_node *cnst = new_Const(mode, tv);
3633 dbg_info *dbg = get_irn_dbg_info(a);
3634 ir_graph *irg = current_ir_graph;
3635 ir_node *block = get_nodes_block(a);
3636 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3637 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3644 } /* transform_node_Minus */
3647 * Transform a Cast_type(Const) into a new Const_type
3649 static ir_node *transform_node_Cast(ir_node *n) {
3651 ir_node *pred = get_Cast_op(n);
3652 ir_type *tp = get_irn_type(n);
3654 if (is_Const(pred) && get_Const_type(pred) != tp) {
3655 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3656 get_Const_tarval(pred), tp);
3657 DBG_OPT_CSTEVAL(oldn, n);
3658 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3659 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3660 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3661 DBG_OPT_CSTEVAL(oldn, n);
3665 } /* transform_node_Cast */
3668 * Transform a Proj(Load) with a non-null address.
3670 static ir_node *transform_node_Proj_Load(ir_node *proj) {
3671 if (get_opt_ldst_only_null_ptr_exceptions()) {
3672 if (get_irn_mode(proj) == mode_X) {
3673 ir_node *load = get_Proj_pred(proj);
3675 /* get the Load address */
3676 const ir_node *addr = get_Load_ptr(load);
3677 const ir_node *confirm;
3679 if (value_not_null(addr, &confirm)) {
3680 if (confirm == NULL) {
3681 /* this node may float if it did not depend on a Confirm */
3682 set_irn_pinned(load, op_pin_state_floats);
3684 if (get_Proj_proj(proj) == pn_Load_X_except) {
3685 DBG_OPT_EXC_REM(proj);
3686 return get_irg_bad(current_ir_graph);
3688 ir_node *blk = get_nodes_block(load);
3689 return new_r_Jmp(current_ir_graph, blk);
3695 } /* transform_node_Proj_Load */
3698 * Transform a Proj(Store) with a non-null address.
3700 static ir_node *transform_node_Proj_Store(ir_node *proj) {
3701 if (get_opt_ldst_only_null_ptr_exceptions()) {
3702 if (get_irn_mode(proj) == mode_X) {
3703 ir_node *store = get_Proj_pred(proj);
3705 /* get the load/store address */
3706 const ir_node *addr = get_Store_ptr(store);
3707 const ir_node *confirm;
3709 if (value_not_null(addr, &confirm)) {
3710 if (confirm == NULL) {
3711 /* this node may float if it did not depend on a Confirm */
3712 set_irn_pinned(store, op_pin_state_floats);
3714 if (get_Proj_proj(proj) == pn_Store_X_except) {
3715 DBG_OPT_EXC_REM(proj);
3716 return get_irg_bad(current_ir_graph);
3718 ir_node *blk = get_nodes_block(store);
3719 return new_r_Jmp(current_ir_graph, blk);
3725 } /* transform_node_Proj_Store */
3728 * Transform a Proj(Div) with a non-zero value.
3729 * Removes the exceptions and routes the memory to the NoMem node.
3731 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3732 ir_node *div = get_Proj_pred(proj);
3733 ir_node *b = get_Div_right(div);
3734 ir_node *res, *new_mem;
3735 const ir_node *confirm;
3738 if (value_not_zero(b, &confirm)) {
3739 /* div(x, y) && y != 0 */
3740 if (confirm == NULL) {
3741 /* we are sure we have a Const != 0 */
3742 new_mem = get_Div_mem(div);
3743 if (is_Pin(new_mem))
3744 new_mem = get_Pin_op(new_mem);
3745 set_Div_mem(div, new_mem);
3746 set_irn_pinned(div, op_pin_state_floats);
3749 proj_nr = get_Proj_proj(proj);
3751 case pn_Div_X_regular:
3752 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3754 case pn_Div_X_except:
3755 /* we found an exception handler, remove it */
3756 DBG_OPT_EXC_REM(proj);
3760 res = get_Div_mem(div);
3761 new_mem = get_irg_no_mem(current_ir_graph);
3764 /* This node can only float up to the Confirm block */
3765 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3767 set_irn_pinned(div, op_pin_state_floats);
3768 /* this is a Div without exception, we can remove the memory edge */
3769 set_Div_mem(div, new_mem);
3774 } /* transform_node_Proj_Div */
3777 * Transform a Proj(Mod) with a non-zero value.
3778 * Removes the exceptions and routes the memory to the NoMem node.
3780 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3781 ir_node *mod = get_Proj_pred(proj);
3782 ir_node *b = get_Mod_right(mod);
3783 ir_node *res, *new_mem;
3784 const ir_node *confirm;
3787 if (value_not_zero(b, &confirm)) {
3788 /* mod(x, y) && y != 0 */
3789 proj_nr = get_Proj_proj(proj);
3791 if (confirm == NULL) {
3792 /* we are sure we have a Const != 0 */
3793 new_mem = get_Mod_mem(mod);
3794 if (is_Pin(new_mem))
3795 new_mem = get_Pin_op(new_mem);
3796 set_Mod_mem(mod, new_mem);
3797 set_irn_pinned(mod, op_pin_state_floats);
3802 case pn_Mod_X_regular:
3803 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3805 case pn_Mod_X_except:
3806 /* we found an exception handler, remove it */
3807 DBG_OPT_EXC_REM(proj);
3811 res = get_Mod_mem(mod);
3812 new_mem = get_irg_no_mem(current_ir_graph);
3815 /* This node can only float up to the Confirm block */
3816 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3818 /* this is a Mod without exception, we can remove the memory edge */
3819 set_Mod_mem(mod, new_mem);
3822 if (get_Mod_left(mod) == b) {
3823 /* a % a = 0 if a != 0 */
3824 ir_mode *mode = get_irn_mode(proj);
3825 ir_node *res = new_Const(mode, get_mode_null(mode));
3827 DBG_OPT_CSTEVAL(mod, res);
3833 } /* transform_node_Proj_Mod */
3836 * Transform a Proj(DivMod) with a non-zero value.
3837 * Removes the exceptions and routes the memory to the NoMem node.
3839 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3840 ir_node *divmod = get_Proj_pred(proj);
3841 ir_node *b = get_DivMod_right(divmod);
3842 ir_node *res, *new_mem;
3843 const ir_node *confirm;
3846 if (value_not_zero(b, &confirm)) {
3847 /* DivMod(x, y) && y != 0 */
3848 proj_nr = get_Proj_proj(proj);
3850 if (confirm == NULL) {
3851 /* we are sure we have a Const != 0 */
3852 new_mem = get_DivMod_mem(divmod);
3853 if (is_Pin(new_mem))
3854 new_mem = get_Pin_op(new_mem);
3855 set_DivMod_mem(divmod, new_mem);
3856 set_irn_pinned(divmod, op_pin_state_floats);
3861 case pn_DivMod_X_regular:
3862 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3864 case pn_DivMod_X_except:
3865 /* we found an exception handler, remove it */
3866 DBG_OPT_EXC_REM(proj);
3870 res = get_DivMod_mem(divmod);
3871 new_mem = get_irg_no_mem(current_ir_graph);
3874 /* This node can only float up to the Confirm block */
3875 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3877 /* this is a DivMod without exception, we can remove the memory edge */
3878 set_DivMod_mem(divmod, new_mem);
3881 case pn_DivMod_res_mod:
3882 if (get_DivMod_left(divmod) == b) {
3883 /* a % a = 0 if a != 0 */
3884 ir_mode *mode = get_irn_mode(proj);
3885 ir_node *res = new_Const(mode, get_mode_null(mode));
3887 DBG_OPT_CSTEVAL(divmod, res);
3893 } /* transform_node_Proj_DivMod */
3896 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3898 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3899 if (get_opt_unreachable_code()) {
3900 ir_node *n = get_Proj_pred(proj);
3901 ir_node *b = get_Cond_selector(n);
3903 if (mode_is_int(get_irn_mode(b))) {
3904 tarval *tb = value_of(b);
3906 if (tb != tarval_bad) {
3907 /* we have a constant switch */
3908 long num = get_Proj_proj(proj);
3910 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3911 if (get_tarval_long(tb) == num) {
3912 /* Do NOT create a jump here, or we will have 2 control flow ops
3913 * in a block. This case is optimized away in optimize_cf(). */
3916 /* this case will NEVER be taken, kill it */
3917 return get_irg_bad(current_ir_graph);
3924 } /* transform_node_Proj_Cond */
3927 * Create a 0 constant of given mode.
3929 static ir_node *create_zero_const(ir_mode *mode) {
3930 tarval *tv = get_mode_null(mode);
3931 ir_node *cnst = new_Const(mode, tv);
3936 /* the order of the values is important! */
3937 typedef enum const_class {
3943 static const_class classify_const(const ir_node* n)
3945 if (is_Const(n)) return const_const;
3946 if (is_irn_constlike(n)) return const_like;
3951 * Determines whether r is more constlike or has a larger index (in that order)
3954 static int operands_are_normalized(const ir_node *l, const ir_node *r)
3956 const const_class l_order = classify_const(l);
3957 const const_class r_order = classify_const(r);
3959 l_order > r_order ||
3960 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3964 * Normalizes and optimizes Cmp nodes.
3966 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
3967 ir_node *n = get_Proj_pred(proj);
3968 ir_node *left = get_Cmp_left(n);
3969 ir_node *right = get_Cmp_right(n);
3972 ir_mode *mode = NULL;
3973 long proj_nr = get_Proj_proj(proj);
3975 /* we can evaluate some cases directly */
3978 return new_Const(mode_b, get_tarval_b_false());
3980 return new_Const(mode_b, get_tarval_b_true());
3982 if (!mode_is_float(get_irn_mode(left)))
3983 return new_Const(mode_b, get_tarval_b_true());
3989 /* remove Casts of both sides */
3990 left = skip_Cast(left);
3991 right = skip_Cast(right);
3993 /* Remove unnecessary conversions */
3994 /* TODO handle constants */
3995 if (is_Conv(left) && is_Conv(right)) {
3996 ir_mode *mode = get_irn_mode(left);
3997 ir_node *op_left = get_Conv_op(left);
3998 ir_node *op_right = get_Conv_op(right);
3999 ir_mode *mode_left = get_irn_mode(op_left);
4000 ir_mode *mode_right = get_irn_mode(op_right);
4002 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
4003 && mode_left != mode_b && mode_right != mode_b) {
4004 ir_graph *irg = current_ir_graph;
4005 ir_node *block = get_nodes_block(n);
4007 if (mode_left == mode_right) {
4011 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
4012 } else if (smaller_mode(mode_left, mode_right)) {
4013 left = new_r_Conv(irg, block, op_left, mode_right);
4016 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4017 } else if (smaller_mode(mode_right, mode_left)) {
4019 right = new_r_Conv(irg, block, op_right, mode_left);
4021 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4026 /* remove operation on both sides if possible */
4027 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4029 * The following operations are NOT safe for floating point operations, for instance
4030 * 1.0 + inf == 2.0 + inf, =/=> x == y
4032 if (mode_is_int(get_irn_mode(left))) {
4033 unsigned lop = get_irn_opcode(left);
4035 if (lop == get_irn_opcode(right)) {
4036 ir_node *ll, *lr, *rl, *rr;
4038 /* same operation on both sides, try to remove */
4042 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
4043 left = get_unop_op(left);
4044 right = get_unop_op(right);
4046 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4049 ll = get_Add_left(left);
4050 lr = get_Add_right(left);
4051 rl = get_Add_left(right);
4052 rr = get_Add_right(right);
4055 /* X + a CMP X + b ==> a CMP b */
4059 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4060 } else if (ll == rr) {
4061 /* X + a CMP b + X ==> a CMP b */
4065 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4066 } else if (lr == rl) {
4067 /* a + X CMP X + b ==> a CMP b */
4071 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4072 } else if (lr == rr) {
4073 /* a + X CMP b + X ==> a CMP b */
4077 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4081 ll = get_Sub_left(left);
4082 lr = get_Sub_right(left);
4083 rl = get_Sub_left(right);
4084 rr = get_Sub_right(right);
4087 /* X - a CMP X - b ==> a CMP b */
4091 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4092 } else if (lr == rr) {
4093 /* a - X CMP b - X ==> a CMP b */
4097 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4101 if (get_Rotl_right(left) == get_Rotl_right(right)) {
4102 /* a ROTL X CMP b ROTL X ==> a CMP b */
4103 left = get_Rotl_left(left);
4104 right = get_Rotl_left(right);
4106 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4114 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4115 if (is_Add(left) || is_Sub(left)) {
4116 ir_node *ll = get_binop_left(left);
4117 ir_node *lr = get_binop_right(left);
4119 if (lr == right && is_Add(left)) {
4126 right = create_zero_const(get_irn_mode(left));
4128 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4131 if (is_Add(right) || is_Sub(right)) {
4132 ir_node *rl = get_binop_left(right);
4133 ir_node *rr = get_binop_right(right);
4135 if (rr == left && is_Add(right)) {
4142 right = create_zero_const(get_irn_mode(left));
4144 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4147 } /* mode_is_int(...) */
4148 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
4150 /* replace mode_b compares with ands/ors */
4151 if (get_irn_mode(left) == mode_b) {
4152 ir_graph *irg = current_ir_graph;
4153 ir_node *block = get_nodes_block(n);
4157 case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4158 case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4159 case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4160 case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4161 case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break;
4162 case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break;
4163 default: bres = NULL;
4166 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4172 * First step: normalize the compare op
4173 * by placing the constant on the right side
4174 * or moving the lower address node to the left.
4176 if (!operands_are_normalized(left, right)) {
4182 proj_nr = get_inversed_pnc(proj_nr);
4187 * Second step: Try to reduce the magnitude
4188 * of a constant. This may help to generate better code
4189 * later and may help to normalize more compares.
4190 * Of course this is only possible for integer values.
4192 if (is_Const(right)) {
4193 mode = get_irn_mode(right);
4194 tv = get_Const_tarval(right);
4196 /* TODO extend to arbitrary constants */
4197 if (is_Conv(left) && tarval_is_null(tv)) {
4198 ir_node *op = get_Conv_op(left);
4199 ir_mode *op_mode = get_irn_mode(op);
4202 * UpConv(x) REL 0 ==> x REL 0
4204 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4205 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
4206 mode_is_signed(mode) || !mode_is_signed(op_mode))) {
4207 tv = get_mode_null(op_mode);
4211 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4215 if (tv != tarval_bad) {
4216 /* the following optimization is possible on modes without Overflow
4217 * on Unary Minus or on == and !=:
4218 * -a CMP c ==> a swap(CMP) -c
4220 * Beware: for two-complement Overflow may occur, so only == and != can
4221 * be optimized, see this:
4222 * -MININT < 0 =/=> MININT > 0 !!!
4224 if (is_Minus(left) &&
4225 (!mode_overflow_on_unary_Minus(mode) ||
4226 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
4227 tv = tarval_neg(tv);
4229 if (tv != tarval_bad) {
4230 left = get_Minus_op(left);
4231 proj_nr = get_inversed_pnc(proj_nr);
4233 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4235 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
4236 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4237 tv = tarval_not(tv);
4239 if (tv != tarval_bad) {
4240 left = get_Not_op(left);
4242 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4246 /* for integer modes, we have more */
4247 if (mode_is_int(mode)) {
4248 /* Ne includes Unordered which is not possible on integers.
4249 * However, frontends often use this wrong, so fix it here */
4250 if (proj_nr & pn_Cmp_Uo) {
4251 proj_nr &= ~pn_Cmp_Uo;
4252 set_Proj_proj(proj, proj_nr);
4255 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4256 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
4257 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
4258 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4260 if (tv != tarval_bad) {
4261 proj_nr ^= pn_Cmp_Eq;
4263 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4266 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4267 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
4268 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
4269 tv = tarval_add(tv, get_mode_one(mode));
4271 if (tv != tarval_bad) {
4272 proj_nr ^= pn_Cmp_Eq;
4274 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4278 /* the following reassociations work only for == and != */
4279 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4281 #if 0 /* Might be not that good in general */
4282 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4283 if (tarval_is_null(tv) && is_Sub(left)) {
4284 right = get_Sub_right(left);
4285 left = get_Sub_left(left);
4287 tv = value_of(right);
4289 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4293 if (tv != tarval_bad) {
4294 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4296 ir_node *c1 = get_Sub_right(left);
4297 tarval *tv2 = value_of(c1);
4299 if (tv2 != tarval_bad) {
4300 tv2 = tarval_add(tv, value_of(c1));
4302 if (tv2 != tarval_bad) {
4303 left = get_Sub_left(left);
4306 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4310 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4311 else if (is_Add(left)) {
4312 ir_node *a_l = get_Add_left(left);
4313 ir_node *a_r = get_Add_right(left);
4317 if (is_Const(a_l)) {
4319 tv2 = value_of(a_l);
4322 tv2 = value_of(a_r);
4325 if (tv2 != tarval_bad) {
4326 tv2 = tarval_sub(tv, tv2, NULL);
4328 if (tv2 != tarval_bad) {
4332 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4336 /* -a == c ==> a == -c, -a != c ==> a != -c */
4337 else if (is_Minus(left)) {
4338 tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4340 if (tv2 != tarval_bad) {
4341 left = get_Minus_op(left);
4344 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4349 /* the following reassociations work only for <= */
4350 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4351 if (tv != tarval_bad) {
4352 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4353 if (get_irn_op(left) == op_Abs) { // TODO something is missing here
4359 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4360 switch (get_irn_opcode(left)) {
4364 c1 = get_And_right(left);
4367 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4368 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4370 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4372 /* TODO: move to constant evaluation */
4373 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4374 c1 = new_Const(mode_b, tv);
4375 DBG_OPT_CSTEVAL(proj, c1);
4379 if (tarval_is_single_bit(tv)) {
4381 * optimization for AND:
4383 * And(x, C) == C ==> And(x, C) != 0
4384 * And(x, C) != C ==> And(X, C) == 0
4386 * if C is a single Bit constant.
4389 /* check for Constant's match. We have check hare the tarvals,
4390 because our const might be changed */
4391 if (get_Const_tarval(c1) == tv) {
4392 /* fine: do the transformation */
4393 tv = get_mode_null(get_tarval_mode(tv));
4394 proj_nr ^= pn_Cmp_Leg;
4396 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4402 c1 = get_Or_right(left);
4403 if (is_Const(c1) && tarval_is_null(tv)) {
4405 * Or(x, C) == 0 && C != 0 ==> FALSE
4406 * Or(x, C) != 0 && C != 0 ==> TRUE
4408 if (! tarval_is_null(get_Const_tarval(c1))) {
4409 /* TODO: move to constant evaluation */
4410 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4411 c1 = new_Const(mode_b, tv);
4412 DBG_OPT_CSTEVAL(proj, c1);
4419 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4421 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4424 c1 = get_Shl_right(left);
4426 tarval *tv1 = get_Const_tarval(c1);
4427 ir_mode *mode = get_irn_mode(left);
4428 tarval *minus1 = get_mode_all_one(mode);
4429 tarval *amask = tarval_shr(minus1, tv1);
4430 tarval *cmask = tarval_shl(minus1, tv1);
4433 if (tarval_and(tv, cmask) != tv) {
4434 /* condition not met */
4435 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4436 c1 = new_Const(mode_b, tv);
4437 DBG_OPT_CSTEVAL(proj, c1);
4440 sl = get_Shl_left(left);
4441 blk = get_nodes_block(n);
4442 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4443 tv = tarval_shr(tv, tv1);
4445 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4450 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4452 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4455 c1 = get_Shr_right(left);
4457 tarval *tv1 = get_Const_tarval(c1);
4458 ir_mode *mode = get_irn_mode(left);
4459 tarval *minus1 = get_mode_all_one(mode);
4460 tarval *amask = tarval_shl(minus1, tv1);
4461 tarval *cmask = tarval_shr(minus1, tv1);
4464 if (tarval_and(tv, cmask) != tv) {
4465 /* condition not met */
4466 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4467 c1 = new_Const(mode_b, tv);
4468 DBG_OPT_CSTEVAL(proj, c1);
4471 sl = get_Shr_left(left);
4472 blk = get_nodes_block(n);
4473 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4474 tv = tarval_shl(tv, tv1);
4476 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4481 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4483 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4486 c1 = get_Shrs_right(left);
4488 tarval *tv1 = get_Const_tarval(c1);
4489 ir_mode *mode = get_irn_mode(left);
4490 tarval *minus1 = get_mode_all_one(mode);
4491 tarval *amask = tarval_shl(minus1, tv1);
4492 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4495 cond = tarval_sub(cond, tv1, NULL);
4496 cond = tarval_shrs(tv, cond);
4498 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4499 /* condition not met */
4500 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4501 c1 = new_Const(mode_b, tv);
4502 DBG_OPT_CSTEVAL(proj, c1);
4505 sl = get_Shrs_left(left);
4506 blk = get_nodes_block(n);
4507 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4508 tv = tarval_shl(tv, tv1);
4510 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4515 } /* tarval != bad */
4518 if (changed & 2) /* need a new Const */
4519 right = new_Const(mode, tv);
4521 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4522 ir_node *op = get_Proj_pred(left);
4524 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4525 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4526 ir_node *c = get_binop_right(op);
4529 tarval *tv = get_Const_tarval(c);
4531 if (tarval_is_single_bit(tv)) {
4532 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4533 ir_node *v = get_binop_left(op);
4534 ir_node *blk = get_irn_n(op, -1);
4535 ir_mode *mode = get_irn_mode(v);
4537 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4538 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode);
4540 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4547 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
4549 /* create a new compare */
4550 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
4551 proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr);
4555 } /* transform_node_Proj_Cmp */
4558 * Optimize CopyB(mem, x, x) into a Nop.
4560 static ir_node *transform_node_Proj_CopyB(ir_node *proj) {
4561 ir_node *copyb = get_Proj_pred(proj);
4562 ir_node *a = get_CopyB_dst(copyb);
4563 ir_node *b = get_CopyB_src(copyb);
4566 switch (get_Proj_proj(proj)) {
4567 case pn_CopyB_X_regular:
4568 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4569 DBG_OPT_EXC_REM(proj);
4570 proj = new_r_Jmp(current_ir_graph, get_nodes_block(copyb));
4572 case pn_CopyB_M_except:
4573 case pn_CopyB_X_except:
4574 DBG_OPT_EXC_REM(proj);
4575 proj = get_irg_bad(current_ir_graph);
4582 } /* transform_node_Proj_CopyB */
4585 * Optimize Bounds(idx, idx, upper) into idx.
4587 static ir_node *transform_node_Proj_Bound(ir_node *proj) {
4588 ir_node *oldn = proj;
4589 ir_node *bound = get_Proj_pred(proj);
4590 ir_node *idx = get_Bound_index(bound);
4591 ir_node *pred = skip_Proj(idx);
4594 if (idx == get_Bound_lower(bound))
4596 else if (is_Bound(pred)) {
4598 * idx was Bounds checked in the same MacroBlock previously,
4599 * it is still valid if lower <= pred_lower && pred_upper <= upper.
4601 ir_node *lower = get_Bound_lower(bound);
4602 ir_node *upper = get_Bound_upper(bound);
4603 if (get_Bound_lower(pred) == lower &&
4604 get_Bound_upper(pred) == upper &&
4605 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
4607 * One could expect that we simply return the previous
4608 * Bound here. However, this would be wrong, as we could
4609 * add an exception Proj to a new location then.
4610 * So, we must turn in into a tuple.
4616 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4617 switch (get_Proj_proj(proj)) {
4619 DBG_OPT_EXC_REM(proj);
4620 proj = get_Bound_mem(bound);
4622 case pn_Bound_X_except:
4623 DBG_OPT_EXC_REM(proj);
4624 proj = get_irg_bad(current_ir_graph);
4628 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4630 case pn_Bound_X_regular:
4631 DBG_OPT_EXC_REM(proj);
4632 proj = new_r_Jmp(current_ir_graph, get_nodes_block(bound));
4639 } /* transform_node_Proj_Bound */
4642 * Does all optimizations on nodes that must be done on it's Proj's
4643 * because of creating new nodes.
4645 static ir_node *transform_node_Proj(ir_node *proj) {
4646 ir_node *n = get_Proj_pred(proj);
4648 if (n->op->ops.transform_node_Proj)
4649 return n->op->ops.transform_node_Proj(proj);
4651 } /* transform_node_Proj */
4654 * Move Confirms down through Phi nodes.
4656 static ir_node *transform_node_Phi(ir_node *phi) {
4658 ir_mode *mode = get_irn_mode(phi);
4660 if (mode_is_reference(mode)) {
4661 n = get_irn_arity(phi);
4663 /* Beware of Phi0 */
4665 ir_node *pred = get_irn_n(phi, 0);
4666 ir_node *bound, *new_Phi, *block, **in;
4669 if (! is_Confirm(pred))
4672 bound = get_Confirm_bound(pred);
4673 pnc = get_Confirm_cmp(pred);
4675 NEW_ARR_A(ir_node *, in, n);
4676 in[0] = get_Confirm_value(pred);
4678 for (i = 1; i < n; ++i) {
4679 pred = get_irn_n(phi, i);
4681 if (! is_Confirm(pred) ||
4682 get_Confirm_bound(pred) != bound ||
4683 get_Confirm_cmp(pred) != pnc)
4685 in[i] = get_Confirm_value(pred);
4687 /* move the Confirm nodes "behind" the Phi */
4688 block = get_irn_n(phi, -1);
4689 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4690 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4694 } /* transform_node_Phi */
4697 * Returns the operands of a commutative bin-op, if one operand is
4698 * a const, it is returned as the second one.
4700 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4701 ir_node *op_a = get_binop_left(binop);
4702 ir_node *op_b = get_binop_right(binop);
4704 assert(is_op_commutative(get_irn_op(binop)));
4706 if (is_Const(op_a)) {
4713 } /* get_comm_Binop_Ops */
4716 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4717 * Such pattern may arise in bitfield stores.
4719 * value c4 value c4 & c2
4720 * AND c3 AND c1 | c3
4727 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4730 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4733 ir_node *and_l, *c3;
4734 ir_node *value, *c4;
4735 ir_node *new_and, *new_const, *block;
4736 ir_mode *mode = get_irn_mode(or);
4738 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4741 get_comm_Binop_Ops(or, &and, &c1);
4742 if (!is_Const(c1) || !is_And(and))
4745 get_comm_Binop_Ops(and, &or_l, &c2);
4749 tv1 = get_Const_tarval(c1);
4750 tv2 = get_Const_tarval(c2);
4752 tv = tarval_or(tv1, tv2);
4753 if (tarval_is_all_one(tv)) {
4754 /* the AND does NOT clear a bit with isn't set by the OR */
4755 set_Or_left(or, or_l);
4756 set_Or_right(or, c1);
4758 /* check for more */
4765 get_comm_Binop_Ops(or_l, &and_l, &c3);
4766 if (!is_Const(c3) || !is_And(and_l))
4769 get_comm_Binop_Ops(and_l, &value, &c4);
4773 /* ok, found the pattern, check for conditions */
4774 assert(mode == get_irn_mode(and));
4775 assert(mode == get_irn_mode(or_l));
4776 assert(mode == get_irn_mode(and_l));
4778 tv3 = get_Const_tarval(c3);
4779 tv4 = get_Const_tarval(c4);
4781 tv = tarval_or(tv4, tv2);
4782 if (!tarval_is_all_one(tv)) {
4783 /* have at least one 0 at the same bit position */
4787 n_tv4 = tarval_not(tv4);
4788 if (tv3 != tarval_and(tv3, n_tv4)) {
4789 /* bit in the or_mask is outside the and_mask */
4793 n_tv2 = tarval_not(tv2);
4794 if (tv1 != tarval_and(tv1, n_tv2)) {
4795 /* bit in the or_mask is outside the and_mask */
4799 /* ok, all conditions met */
4800 block = get_irn_n(or, -1);
4802 new_and = new_r_And(current_ir_graph, block,
4803 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
4805 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
4807 set_Or_left(or, new_and);
4808 set_Or_right(or, new_const);
4810 /* check for more */
4812 } /* transform_node_Or_bf_store */
4815 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4817 static ir_node *transform_node_Or_Rotl(ir_node *or) {
4818 ir_mode *mode = get_irn_mode(or);
4819 ir_node *shl, *shr, *block;
4820 ir_node *irn, *x, *c1, *c2, *v, *sub, *n, *rotval;
4823 if (! mode_is_int(mode))
4826 shl = get_binop_left(or);
4827 shr = get_binop_right(or);
4836 } else if (!is_Shl(shl)) {
4838 } else if (!is_Shr(shr)) {
4841 x = get_Shl_left(shl);
4842 if (x != get_Shr_left(shr))
4845 c1 = get_Shl_right(shl);
4846 c2 = get_Shr_right(shr);
4847 if (is_Const(c1) && is_Const(c2)) {
4848 tv1 = get_Const_tarval(c1);
4849 if (! tarval_is_long(tv1))
4852 tv2 = get_Const_tarval(c2);
4853 if (! tarval_is_long(tv2))
4856 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4857 != (int) get_mode_size_bits(mode))
4860 /* yet, condition met */
4861 block = get_nodes_block(or);
4863 n = new_r_Rotl(current_ir_graph, block, x, c1, mode);
4865 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4872 rotval = sub; /* a Rot right is not supported, so use a rot left */
4873 } else if (is_Sub(c2)) {
4879 if (get_Sub_right(sub) != v)
4882 c1 = get_Sub_left(sub);
4886 tv1 = get_Const_tarval(c1);
4887 if (! tarval_is_long(tv1))
4890 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4893 /* yet, condition met */
4894 block = get_nodes_block(or);
4896 n = new_r_Rotl(current_ir_graph, block, x, rotval, mode);
4898 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL);
4900 } /* transform_node_Or_Rotl */
4905 static ir_node *transform_node_Or(ir_node *n) {
4906 ir_node *c, *oldn = n;
4907 ir_node *a = get_Or_left(n);
4908 ir_node *b = get_Or_right(n);
4911 if (is_Not(a) && is_Not(b)) {
4912 /* ~a | ~b = ~(a&b) */
4913 ir_node *block = get_nodes_block(n);
4915 mode = get_irn_mode(n);
4918 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4919 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4920 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4924 /* we can evaluate 2 Projs of the same Cmp */
4925 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4926 ir_node *pred_a = get_Proj_pred(a);
4927 ir_node *pred_b = get_Proj_pred(b);
4928 if (pred_a == pred_b) {
4929 dbg_info *dbgi = get_irn_dbg_info(n);
4930 ir_node *block = get_nodes_block(pred_a);
4931 pn_Cmp pn_a = get_Proj_proj(a);
4932 pn_Cmp pn_b = get_Proj_proj(b);
4933 /* yes, we can simply calculate with pncs */
4934 pn_Cmp new_pnc = pn_a | pn_b;
4936 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
4941 mode = get_irn_mode(n);
4942 HANDLE_BINOP_PHI(tarval_or, a, b, c, mode);
4944 n = transform_node_Or_bf_store(n);
4945 n = transform_node_Or_Rotl(n);
4949 n = transform_bitwise_distributive(n, transform_node_Or);
4952 } /* transform_node_Or */
4956 static ir_node *transform_node(ir_node *n);
4959 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
4961 * Should be moved to reassociation?
4963 static ir_node *transform_node_shift(ir_node *n) {
4964 ir_node *left, *right;
4966 tarval *tv1, *tv2, *res;
4967 ir_node *in[2], *irn, *block;
4969 left = get_binop_left(n);
4971 /* different operations */
4972 if (get_irn_op(left) != get_irn_op(n))
4975 right = get_binop_right(n);
4976 tv1 = value_of(right);
4977 if (tv1 == tarval_bad)
4980 tv2 = value_of(get_binop_right(left));
4981 if (tv2 == tarval_bad)
4984 res = tarval_add(tv1, tv2);
4985 mode = get_irn_mode(n);
4987 /* beware: a simple replacement works only, if res < modulo shift */
4989 int modulo_shf = get_mode_modulo_shift(mode);
4990 assert(modulo_shf >= (int) get_mode_size_bits(mode));
4991 if (modulo_shf > 0) {
4992 tarval *modulo = new_tarval_from_long(modulo_shf,
4993 get_tarval_mode(res));
4995 /* shifting too much */
4996 if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) {
4998 ir_graph *irg = get_irn_irg(n);
4999 ir_node *block = get_nodes_block(n);
5000 dbg_info *dbgi = get_irn_dbg_info(n);
5001 ir_node *cnst = new_Const(mode_Iu, new_tarval_from_long(get_mode_size_bits(mode)-1, mode_Iu));
5002 return new_rd_Shrs(dbgi, irg, block, get_binop_left(left),
5006 return new_Const(mode, get_mode_null(mode));
5010 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
5013 /* ok, we can replace it */
5014 block = get_nodes_block(n);
5016 in[0] = get_binop_left(left);
5017 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
5019 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
5021 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5023 return transform_node(irn);
5024 } /* transform_node_shift */
5027 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
5029 * - and, or, xor instead of &
5030 * - Shl, Shr, Shrs, rotl instead of >>
5031 * (with a special case for Or/Xor + Shrs)
5033 static ir_node *transform_node_bitop_shift(ir_node *n) {
5035 ir_node *right = get_binop_right(n);
5036 ir_mode *mode = get_irn_mode(n);
5037 ir_node *bitop_left;
5038 ir_node *bitop_right;
5050 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
5052 if (!is_Const(right))
5055 left = get_binop_left(n);
5056 op_left = get_irn_op(left);
5057 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
5060 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
5061 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
5062 /* TODO: test if sign bit is affectes */
5066 bitop_right = get_binop_right(left);
5067 if (!is_Const(bitop_right))
5070 bitop_left = get_binop_left(left);
5072 irg = get_irn_irg(n);
5073 block = get_nodes_block(n);
5074 dbgi = get_irn_dbg_info(n);
5075 tv1 = get_Const_tarval(bitop_right);
5076 tv2 = get_Const_tarval(right);
5078 assert(get_tarval_mode(tv1) == mode);
5081 new_shift = new_rd_Shl(dbgi, irg, block, bitop_left, right, mode);
5082 tv_shift = tarval_shl(tv1, tv2);
5083 } else if(is_Shr(n)) {
5084 new_shift = new_rd_Shr(dbgi, irg, block, bitop_left, right, mode);
5085 tv_shift = tarval_shr(tv1, tv2);
5086 } else if(is_Shrs(n)) {
5087 new_shift = new_rd_Shrs(dbgi, irg, block, bitop_left, right, mode);
5088 tv_shift = tarval_shrs(tv1, tv2);
5091 new_shift = new_rd_Rotl(dbgi, irg, block, bitop_left, right, mode);
5092 tv_shift = tarval_rotl(tv1, tv2);
5095 assert(get_tarval_mode(tv_shift) == mode);
5096 new_const = new_Const(mode, tv_shift);
5098 if (op_left == op_And) {
5099 new_bitop = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5100 } else if(op_left == op_Or) {
5101 new_bitop = new_rd_Or(dbgi, irg, block, new_shift, new_const, mode);
5103 assert(op_left == op_Eor);
5104 new_bitop = new_rd_Eor(dbgi, irg, block, new_shift, new_const, mode);
5112 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5114 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5115 * (also with x >>s c1 when c1>=c2)
5117 static ir_node *transform_node_shl_shr(ir_node *n) {
5119 ir_node *right = get_binop_right(n);
5135 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5137 if (!is_Const(right))
5140 left = get_binop_left(n);
5141 mode = get_irn_mode(n);
5142 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5143 ir_node *shr_right = get_binop_right(left);
5145 if (!is_Const(shr_right))
5148 x = get_binop_left(left);
5149 tv_shr = get_Const_tarval(shr_right);
5150 tv_shl = get_Const_tarval(right);
5152 if (is_Shrs(left)) {
5153 /* shrs variant only allowed if c1 >= c2 */
5154 if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge))
5157 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5160 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5162 tv_mask = tarval_shl(tv_mask, tv_shl);
5163 } else if(is_Shr(n) && is_Shl(left)) {
5164 ir_node *shl_right = get_Shl_right(left);
5166 if (!is_Const(shl_right))
5169 x = get_Shl_left(left);
5170 tv_shr = get_Const_tarval(right);
5171 tv_shl = get_Const_tarval(shl_right);
5173 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5174 tv_mask = tarval_shr(tv_mask, tv_shr);
5179 assert(get_tarval_mode(tv_shl) == get_tarval_mode(tv_shr));
5180 assert(tv_mask != tarval_bad);
5181 assert(get_tarval_mode(tv_mask) == mode);
5183 irg = get_irn_irg(n);
5184 block = get_nodes_block(n);
5185 dbgi = get_irn_dbg_info(n);
5187 pnc = tarval_cmp(tv_shl, tv_shr);
5188 if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) {
5189 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5190 new_const = new_Const(get_tarval_mode(tv_shift), tv_shift);
5192 new_shift = new_rd_Shrs(dbgi, irg, block, x, new_const, mode);
5194 new_shift = new_rd_Shr(dbgi, irg, block, x, new_const, mode);
5197 assert(pnc == pn_Cmp_Gt);
5198 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5199 new_const = new_Const(get_tarval_mode(tv_shift), tv_shift);
5200 new_shift = new_rd_Shl(dbgi, irg, block, x, new_const, mode);
5203 new_const = new_Const(mode, tv_mask);
5204 new_and = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5212 static ir_node *transform_node_Shr(ir_node *n) {
5213 ir_node *c, *oldn = n;
5214 ir_node *left = get_Shr_left(n);
5215 ir_node *right = get_Shr_right(n);
5216 ir_mode *mode = get_irn_mode(n);
5218 HANDLE_BINOP_PHI(tarval_shr, left, right, c, mode);
5219 n = transform_node_shift(n);
5222 n = transform_node_shl_shr(n);
5224 n = transform_node_bitop_shift(n);
5227 } /* transform_node_Shr */
5232 static ir_node *transform_node_Shrs(ir_node *n) {
5233 ir_node *c, *oldn = n;
5234 ir_node *a = get_Shrs_left(n);
5235 ir_node *b = get_Shrs_right(n);
5236 ir_mode *mode = get_irn_mode(n);
5238 HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode);
5239 n = transform_node_shift(n);
5242 n = transform_node_bitop_shift(n);
5245 } /* transform_node_Shrs */
5250 static ir_node *transform_node_Shl(ir_node *n) {
5251 ir_node *c, *oldn = n;
5252 ir_node *a = get_Shl_left(n);
5253 ir_node *b = get_Shl_right(n);
5254 ir_mode *mode = get_irn_mode(n);
5256 HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode);
5257 n = transform_node_shift(n);
5260 n = transform_node_shl_shr(n);
5262 n = transform_node_bitop_shift(n);
5265 } /* transform_node_Shl */
5270 static ir_node *transform_node_Rotl(ir_node *n) {
5271 ir_node *c, *oldn = n;
5272 ir_node *a = get_Rotl_left(n);
5273 ir_node *b = get_Rotl_right(n);
5274 ir_mode *mode = get_irn_mode(n);
5276 HANDLE_BINOP_PHI(tarval_rotl, a, b, c, mode);
5277 n = transform_node_shift(n);
5280 n = transform_node_bitop_shift(n);
5283 } /* transform_node_Rotl */
5288 static ir_node *transform_node_Conv(ir_node *n) {
5289 ir_node *c, *oldn = n;
5290 ir_node *a = get_Conv_op(n);
5292 if (is_const_Phi(a)) {
5293 c = apply_conv_on_phi(a, get_irn_mode(n));
5295 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5300 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5301 ir_mode *mode = get_irn_mode(n);
5302 return new_r_Unknown(current_ir_graph, mode);
5306 } /* transform_node_Conv */
5309 * Remove dead blocks and nodes in dead blocks
5310 * in keep alive list. We do not generate a new End node.
5312 static ir_node *transform_node_End(ir_node *n) {
5313 int i, j, n_keepalives = get_End_n_keepalives(n);
5316 NEW_ARR_A(ir_node *, in, n_keepalives);
5318 for (i = j = 0; i < n_keepalives; ++i) {
5319 ir_node *ka = get_End_keepalive(n, i);
5321 if (! is_Block_dead(ka)) {
5325 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
5328 /* FIXME: beabi need to keep a Proj(M) */
5329 if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka))
5332 if (j != n_keepalives)
5333 set_End_keepalives(n, j, in);
5335 } /* transform_node_End */
5337 /** returns 1 if a == -b */
5338 static int is_negated_value(ir_node *a, ir_node *b) {
5339 if (is_Minus(a) && get_Minus_op(a) == b)
5341 if (is_Minus(b) && get_Minus_op(b) == a)
5343 if (is_Sub(a) && is_Sub(b)) {
5344 ir_node *a_left = get_Sub_left(a);
5345 ir_node *a_right = get_Sub_right(a);
5346 ir_node *b_left = get_Sub_left(b);
5347 ir_node *b_right = get_Sub_right(b);
5349 if (a_left == b_right && a_right == b_left)
5357 * Optimize a Mux into some simpler cases.
5359 static ir_node *transform_node_Mux(ir_node *n) {
5360 ir_node *oldn = n, *sel = get_Mux_sel(n);
5361 ir_mode *mode = get_irn_mode(n);
5362 ir_node *t = get_Mux_true(n);
5363 ir_node *f = get_Mux_false(n);
5364 ir_graph *irg = current_ir_graph;
5365 ir_node *conds[1], *vals[2];
5367 /* first normalization step: move a possible zero to the false case */
5369 ir_node *cmp = get_Proj_pred(sel);
5372 if (is_Const(t) && is_Const_null(t)) {
5373 /* Psi(x, 0, y) => Psi(x, y, 0) */
5374 pn_Cmp pnc = get_Proj_proj(sel);
5375 sel = new_r_Proj(irg, get_nodes_block(cmp), cmp, mode_b,
5376 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
5380 n = new_rd_Psi(get_irn_dbg_info(n), irg, get_nodes_block(n), 1, conds, vals, mode);
5387 /* note: after normalization, false can only happen on default */
5388 if (mode == mode_b) {
5389 dbg_info *dbg = get_irn_dbg_info(n);
5390 ir_node *block = get_nodes_block(n);
5391 ir_graph *irg = current_ir_graph;
5394 tarval *tv_t = get_Const_tarval(t);
5395 if (tv_t == tarval_b_true) {
5397 /* Muxb(sel, true, false) = sel */
5398 assert(get_Const_tarval(f) == tarval_b_false);
5399 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5402 /* Muxb(sel, true, x) = Or(sel, x) */
5403 n = new_rd_Or(dbg, irg, block, sel, f, mode_b);
5404 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5408 } else if (is_Const(f)) {
5409 tarval *tv_f = get_Const_tarval(f);
5410 if (tv_f == tarval_b_true) {
5411 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5412 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
5413 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5414 n = new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
5417 /* Muxb(sel, x, false) = And(sel, x) */
5418 assert(tv_f == tarval_b_false);
5419 n = new_rd_And(dbg, irg, block, sel, t, mode_b);
5420 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5426 /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */
5427 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5428 tarval *a = get_Const_tarval(t);
5429 tarval *b = get_Const_tarval(f);
5430 tarval *null = get_tarval_null(mode);
5433 if (tarval_cmp(a, b) & pn_Cmp_Gt) {
5434 diff = tarval_sub(a, b, NULL);
5437 diff = tarval_sub(b, a, NULL);
5441 if (diff == get_tarval_one(mode) && min != null) {
5442 dbg_info *dbg = get_irn_dbg_info(n);
5443 ir_node *block = get_nodes_block(n);
5444 ir_graph *irg = current_ir_graph;
5448 vals[0] = new_Const(mode, tarval_sub(a, min, NULL));
5449 vals[1] = new_Const(mode, tarval_sub(b, min, NULL));
5450 n = new_rd_Psi(dbg, irg, block, 1, conds, vals, mode);
5451 n = new_rd_Add(dbg, irg, block, n, new_Const(mode, min), mode);
5457 ir_node *cmp = get_Proj_pred(sel);
5458 long pn = get_Proj_proj(sel);
5461 * Note: normalization puts the constant on the right side,
5462 * so we check only one case.
5464 * Note further that these optimization work even for floating point
5465 * with NaN's because -NaN == NaN.
5466 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5470 ir_node *cmp_r = get_Cmp_right(cmp);
5471 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5472 ir_node *block = get_nodes_block(n);
5473 ir_node *cmp_l = get_Cmp_left(cmp);
5475 if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
5478 if ( (cmp_l == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
5479 || (cmp_l == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
5481 /* Psi(a >/>= 0, a, -a) = Psi(a </<= 0, -a, a) ==> Abs(a) */
5482 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5484 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5486 } else if ((cmp_l == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
5487 || (cmp_l == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
5489 /* Psi(a </<= 0, a, -a) = Psi(a >/>= 0, -a, a) ==> -Abs(a) */
5490 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5492 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
5494 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5499 if (mode_is_int(mode)) {
5501 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
5502 /* Psi((a & b) != 0, c, 0) */
5503 ir_node *and_r = get_And_right(cmp_l);
5506 if (and_r == t && f == cmp_r) {
5507 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5508 if (pn == pn_Cmp_Lg) {
5509 /* Psi((a & 2^C) != 0, 2^C, 0) */
5512 /* Psi((a & 2^C) == 0, 2^C, 0) */
5513 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5514 block, cmp_l, t, mode);
5519 if (is_Shl(and_r)) {
5520 ir_node *shl_l = get_Shl_left(and_r);
5521 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5522 if (and_r == t && f == cmp_r) {
5523 if (pn == pn_Cmp_Lg) {
5524 /* (a & (1 << n)) != 0, (1 << n), 0) */
5527 /* (a & (1 << n)) == 0, (1 << n), 0) */
5528 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5529 block, cmp_l, t, mode);
5535 and_l = get_And_left(cmp_l);
5536 if (is_Shl(and_l)) {
5537 ir_node *shl_l = get_Shl_left(and_l);
5538 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5539 if (and_l == t && f == cmp_r) {
5540 if (pn == pn_Cmp_Lg) {
5541 /* ((1 << n) & a) != 0, (1 << n), 0) */
5544 /* ((1 << n) & a) == 0, (1 << n), 0) */
5545 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5546 block, cmp_l, t, mode);
5557 return arch_transform_node_Mux(n);
5558 } /* transform_node_Mux */
5561 * Optimize a Psi into some simpler cases.
5563 static ir_node *transform_node_Psi(ir_node *n) {
5565 return transform_node_Mux(n);
5568 } /* transform_node_Psi */
5571 * optimize Sync nodes that have other syncs as input we simply add the inputs
5572 * of the other sync to our own inputs
5574 static ir_node *transform_node_Sync(ir_node *n) {
5575 int arity = get_Sync_n_preds(n);
5578 for (i = 0; i < arity;) {
5579 ir_node *pred = get_Sync_pred(n, i);
5583 if (!is_Sync(pred)) {
5591 pred_arity = get_Sync_n_preds(pred);
5592 for (j = 0; j < pred_arity; ++j) {
5593 ir_node *pred_pred = get_Sync_pred(pred, j);
5598 add_irn_n(n, pred_pred);
5602 if (get_Sync_pred(n, k) == pred_pred) break;
5607 /* rehash the sync node */
5608 add_identities(current_ir_graph->value_table, n);
5614 * Tries several [inplace] [optimizing] transformations and returns an
5615 * equivalent node. The difference to equivalent_node() is that these
5616 * transformations _do_ generate new nodes, and thus the old node must
5617 * not be freed even if the equivalent node isn't the old one.
5619 static ir_node *transform_node(ir_node *n) {
5623 * Transform_node is the only "optimizing transformation" that might
5624 * return a node with a different opcode. We iterate HERE until fixpoint
5625 * to get the final result.
5629 if (n->op->ops.transform_node)
5630 n = n->op->ops.transform_node(n);
5631 } while (oldn != n);
5634 } /* transform_node */
5637 * Sets the default transform node operation for an ir_op_ops.
5639 * @param code the opcode for the default operation
5640 * @param ops the operations initialized
5645 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5649 ops->transform_node = transform_node_##a; \
5651 #define CASE_PROJ(a) \
5653 ops->transform_node_Proj = transform_node_Proj_##a; \
5655 #define CASE_PROJ_EX(a) \
5657 ops->transform_node = transform_node_##a; \
5658 ops->transform_node_Proj = transform_node_Proj_##a; \
5667 CASE_PROJ_EX(DivMod);
5702 } /* firm_set_default_transform_node */
5705 /* **************** Common Subexpression Elimination **************** */
5707 /** The size of the hash table used, should estimate the number of nodes
5709 #define N_IR_NODES 512
5711 /** Compares the attributes of two Const nodes. */
5712 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
5713 return (get_Const_tarval(a) != get_Const_tarval(b))
5714 || (get_Const_type(a) != get_Const_type(b));
5715 } /* node_cmp_attr_Const */
5717 /** Compares the attributes of two Proj nodes. */
5718 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
5719 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
5720 } /* node_cmp_attr_Proj */
5722 /** Compares the attributes of two Filter nodes. */
5723 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
5724 return get_Filter_proj(a) != get_Filter_proj(b);
5725 } /* node_cmp_attr_Filter */
5727 /** Compares the attributes of two Alloc nodes. */
5728 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
5729 const alloc_attr *pa = get_irn_alloc_attr(a);
5730 const alloc_attr *pb = get_irn_alloc_attr(b);
5731 return (pa->where != pb->where) || (pa->type != pb->type);
5732 } /* node_cmp_attr_Alloc */
5734 /** Compares the attributes of two Free nodes. */
5735 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
5736 const free_attr *pa = get_irn_free_attr(a);
5737 const free_attr *pb = get_irn_free_attr(b);
5738 return (pa->where != pb->where) || (pa->type != pb->type);
5739 } /* node_cmp_attr_Free */
5741 /** Compares the attributes of two SymConst nodes. */
5742 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
5743 const symconst_attr *pa = get_irn_symconst_attr(a);
5744 const symconst_attr *pb = get_irn_symconst_attr(b);
5745 return (pa->kind != pb->kind)
5746 || (pa->sym.type_p != pb->sym.type_p)
5747 || (pa->tp != pb->tp);
5748 } /* node_cmp_attr_SymConst */
5750 /** Compares the attributes of two Call nodes. */
5751 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5752 return get_irn_call_attr(a) != get_irn_call_attr(b);
5753 } /* node_cmp_attr_Call */
5755 /** Compares the attributes of two Sel nodes. */
5756 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
5757 const ir_entity *a_ent = get_Sel_entity(a);
5758 const ir_entity *b_ent = get_Sel_entity(b);
5760 (a_ent->kind != b_ent->kind) ||
5761 (a_ent->name != b_ent->name) ||
5762 (a_ent->owner != b_ent->owner) ||
5763 (a_ent->ld_name != b_ent->ld_name) ||
5764 (a_ent->type != b_ent->type);
5765 } /* node_cmp_attr_Sel */
5767 /** Compares the attributes of two Phi nodes. */
5768 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
5769 /* we can only enter this function if both nodes have the same number of inputs,
5770 hence it is enough to check if one of them is a Phi0 */
5772 /* check the Phi0 pos attribute */
5773 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
5776 } /* node_cmp_attr_Phi */
5778 /** Compares the attributes of two Conv nodes. */
5779 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
5780 return get_Conv_strict(a) != get_Conv_strict(b);
5781 } /* node_cmp_attr_Conv */
5783 /** Compares the attributes of two Cast nodes. */
5784 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
5785 return get_Cast_type(a) != get_Cast_type(b);
5786 } /* node_cmp_attr_Cast */
5788 /** Compares the attributes of two Load nodes. */
5789 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
5790 if (get_Load_volatility(a) == volatility_is_volatile ||
5791 get_Load_volatility(b) == volatility_is_volatile)
5792 /* NEVER do CSE on volatile Loads */
5794 /* do not CSE Loads with different alignment. Be conservative. */
5795 if (get_Load_align(a) != get_Load_align(b))
5798 return get_Load_mode(a) != get_Load_mode(b);
5799 } /* node_cmp_attr_Load */
5801 /** Compares the attributes of two Store nodes. */
5802 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
5803 /* do not CSE Stores with different alignment. Be conservative. */
5804 if (get_Store_align(a) != get_Store_align(b))
5807 /* NEVER do CSE on volatile Stores */
5808 return (get_Store_volatility(a) == volatility_is_volatile ||
5809 get_Store_volatility(b) == volatility_is_volatile);
5810 } /* node_cmp_attr_Store */
5812 /** Compares two exception attributes */
5813 static int node_cmp_exception(ir_node *a, ir_node *b) {
5814 const except_attr *ea = get_irn_except_attr(a);
5815 const except_attr *eb = get_irn_except_attr(b);
5817 return ea->pin_state != eb->pin_state;
5820 #define node_cmp_attr_Bound node_cmp_exception
5822 /** Compares the attributes of two Div nodes. */
5823 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
5824 const divmod_attr *ma = get_irn_divmod_attr(a);
5825 const divmod_attr *mb = get_irn_divmod_attr(b);
5826 return ma->exc.pin_state != mb->exc.pin_state ||
5827 ma->res_mode != mb->res_mode ||
5828 ma->no_remainder != mb->no_remainder;
5829 } /* node_cmp_attr_Div */
5831 /** Compares the attributes of two DivMod nodes. */
5832 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
5833 const divmod_attr *ma = get_irn_divmod_attr(a);
5834 const divmod_attr *mb = get_irn_divmod_attr(b);
5835 return ma->exc.pin_state != mb->exc.pin_state ||
5836 ma->res_mode != mb->res_mode;
5837 } /* node_cmp_attr_DivMod */
5839 /** Compares the attributes of two Mod nodes. */
5840 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
5841 const divmod_attr *ma = get_irn_divmod_attr(a);
5842 const divmod_attr *mb = get_irn_divmod_attr(b);
5843 return ma->exc.pin_state != mb->exc.pin_state ||
5844 ma->res_mode != mb->res_mode;
5845 } /* node_cmp_attr_Mod */
5847 /** Compares the attributes of two Quot nodes. */
5848 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
5849 const divmod_attr *ma = get_irn_divmod_attr(a);
5850 const divmod_attr *mb = get_irn_divmod_attr(b);
5851 return ma->exc.pin_state != mb->exc.pin_state ||
5852 ma->res_mode != mb->res_mode;
5853 } /* node_cmp_attr_Quot */
5855 /** Compares the attributes of two Confirm nodes. */
5856 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
5857 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5858 } /* node_cmp_attr_Confirm */
5860 /** Compares the attributes of two ASM nodes. */
5861 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
5863 const ir_asm_constraint *ca;
5864 const ir_asm_constraint *cb;
5867 if (get_ASM_text(a) != get_ASM_text(b))
5870 /* Should we really check the constraints here? Should be better, but is strange. */
5871 n = get_ASM_n_input_constraints(a);
5872 if (n != get_ASM_n_input_constraints(b))
5875 ca = get_ASM_input_constraints(a);
5876 cb = get_ASM_input_constraints(b);
5877 for (i = 0; i < n; ++i) {
5878 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5882 n = get_ASM_n_output_constraints(a);
5883 if (n != get_ASM_n_output_constraints(b))
5886 ca = get_ASM_output_constraints(a);
5887 cb = get_ASM_output_constraints(b);
5888 for (i = 0; i < n; ++i) {
5889 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5893 n = get_ASM_n_clobbers(a);
5894 if (n != get_ASM_n_clobbers(b))
5897 cla = get_ASM_clobbers(a);
5898 clb = get_ASM_clobbers(b);
5899 for (i = 0; i < n; ++i) {
5900 if (cla[i] != clb[i])
5904 } /* node_cmp_attr_ASM */
5907 * Set the default node attribute compare operation for an ir_op_ops.
5909 * @param code the opcode for the default operation
5910 * @param ops the operations initialized
5915 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
5919 ops->node_cmp_attr = node_cmp_attr_##a; \
5950 } /* firm_set_default_node_cmp_attr */
5953 * Compare function for two nodes in the value table. Gets two
5954 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
5956 int identities_cmp(const void *elt, const void *key) {
5957 ir_node *a = (ir_node *)elt;
5958 ir_node *b = (ir_node *)key;
5961 if (a == b) return 0;
5963 if ((get_irn_op(a) != get_irn_op(b)) ||
5964 (get_irn_mode(a) != get_irn_mode(b))) return 1;
5966 /* compare if a's in and b's in are of equal length */
5967 irn_arity_a = get_irn_intra_arity(a);
5968 if (irn_arity_a != get_irn_intra_arity(b))
5971 if (get_irn_pinned(a) == op_pin_state_pinned) {
5972 /* for pinned nodes, the block inputs must be equal */
5973 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
5975 } else if (! get_opt_global_cse()) {
5976 /* for block-local CSE both nodes must be in the same MacroBlock */
5977 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
5981 /* compare a->in[0..ins] with b->in[0..ins] */
5982 for (i = 0; i < irn_arity_a; i++)
5983 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
5987 * here, we already now that the nodes are identical except their
5990 if (a->op->ops.node_cmp_attr)
5991 return a->op->ops.node_cmp_attr(a, b);
5994 } /* identities_cmp */
5997 * Calculate a hash value of a node.
5999 * @param node The IR-node
6001 unsigned ir_node_hash(const ir_node *node) {
6002 return node->op->ops.hash(node);
6003 } /* ir_node_hash */
6006 pset *new_identities(void) {
6007 return new_pset(identities_cmp, N_IR_NODES);
6008 } /* new_identities */
6010 void del_identities(pset *value_table) {
6011 del_pset(value_table);
6012 } /* del_identities */
6015 * Normalize a node by putting constants (and operands with larger
6016 * node index) on the right (operator side).
6018 * @param n The node to normalize
6020 static void normalize_node(ir_node *n) {
6021 if (is_op_commutative(get_irn_op(n))) {
6022 ir_node *l = get_binop_left(n);
6023 ir_node *r = get_binop_right(n);
6025 /* For commutative operators perform a OP b == b OP a but keep
6026 * constants on the RIGHT side. This helps greatly in some
6027 * optimizations. Moreover we use the idx number to make the form
6029 if (!operands_are_normalized(l, r)) {
6030 set_binop_left(n, r);
6031 set_binop_right(n, l);
6034 } /* normalize_node */
6037 * Update the nodes after a match in the value table. If both nodes have
6038 * the same MacroBlock but different Blocks, we must ensure that the node
6039 * with the dominating Block (the node that is near to the MacroBlock header
6040 * is stored in the table.
6041 * Because a MacroBlock has only one "non-exception" flow, we don't need
6042 * dominance info here: We known, that one block must dominate the other and
6043 * following the only block input will allow to find it.
6045 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
6046 ir_node *known_blk, *new_block, *block, *mbh;
6048 if (get_opt_global_cse()) {
6049 /* Block inputs are meaning less */
6052 known_blk = get_irn_n(known_irn, -1);
6053 new_block = get_irn_n(new_ir_node, -1);
6054 if (known_blk == new_block) {
6055 /* already in the same block */
6059 * We expect the typical case when we built the graph. In that case, the
6060 * known_irn is already the upper one, so checking this should be faster.
6063 mbh = get_Block_MacroBlock(new_block);
6065 if (block == known_blk) {
6066 /* ok, we have found it: known_block dominates new_block as expected */
6071 * We have reached the MacroBlock header NOT founding
6072 * the known_block. new_block must dominate known_block.
6075 set_irn_n(known_irn, -1, new_block);
6078 assert(get_Block_n_cfgpreds(block) == 1);
6079 block = get_Block_cfgpred_block(block, 0);
6081 } /* update_value_table */
6084 * Return the canonical node computing the same value as n.
6085 * Looks up the node in a hash table, enters it in the table
6086 * if it isn't there yet.
6088 * @param value_table the HashSet containing all nodes in the
6090 * @param n the node to look up
6092 * @return a node that computes the same value as n or n if no such
6093 * node could be found
6095 ir_node *identify_remember(pset *value_table, ir_node *n) {
6098 if (!value_table) return n;
6101 /* lookup or insert in hash table with given hash key. */
6102 o = pset_insert(value_table, n, ir_node_hash(n));
6105 update_known_irn(o, n);
6110 } /* identify_remember */
6113 * During construction we set the op_pin_state_pinned flag in the graph right when the
6114 * optimization is performed. The flag turning on procedure global cse could
6115 * be changed between two allocations. This way we are safe.
6117 * @param value_table The value table
6118 * @param n The node to lookup
6120 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
6123 n = identify_remember(value_table, n);
6124 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
6125 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6127 } /* identify_cons */
6129 /* Add a node to the identities value table. */
6130 void add_identities(pset *value_table, ir_node *node) {
6131 if (get_opt_cse() && is_no_Block(node))
6132 identify_remember(value_table, node);
6133 } /* add_identities */
6135 /* Visit each node in the value table of a graph. */
6136 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
6138 ir_graph *rem = current_ir_graph;
6140 current_ir_graph = irg;
6141 foreach_pset(irg->value_table, node)
6143 current_ir_graph = rem;
6144 } /* visit_all_identities */
6147 * Garbage in, garbage out. If a node has a dead input, i.e., the
6148 * Bad node is input to the node, return the Bad node.
6150 static ir_node *gigo(ir_node *node) {
6152 ir_op *op = get_irn_op(node);
6154 /* remove garbage blocks by looking at control flow that leaves the block
6155 and replacing the control flow by Bad. */
6156 if (get_irn_mode(node) == mode_X) {
6157 ir_node *block = get_nodes_block(skip_Proj(node));
6159 /* Don't optimize nodes in immature blocks. */
6160 if (!get_Block_matured(block))
6162 /* Don't optimize End, may have Bads. */
6163 if (op == op_End) return node;
6165 if (is_Block(block)) {
6166 if (is_Block_dead(block)) {
6167 /* control flow from dead block is dead */
6171 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
6172 if (!is_Bad(get_irn_n(block, i)))
6176 ir_graph *irg = get_irn_irg(block);
6177 /* the start block is never dead */
6178 if (block != get_irg_start_block(irg)
6179 && block != get_irg_end_block(irg)) {
6181 * Do NOT kill control flow without setting
6182 * the block to dead of bad things can happen:
6183 * We get a Block that is not reachable be irg_block_walk()
6184 * but can be found by irg_walk()!
6186 set_Block_dead(block);
6193 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6194 blocks predecessors is dead. */
6195 if (op != op_Block && op != op_Phi && op != op_Tuple) {
6196 irn_arity = get_irn_arity(node);
6199 * Beware: we can only read the block of a non-floating node.
6201 if (is_irn_pinned_in_irg(node) &&
6202 is_Block_dead(get_nodes_block(skip_Proj(node))))
6205 for (i = 0; i < irn_arity; i++) {
6206 ir_node *pred = get_irn_n(node, i);
6211 /* Propagating Unknowns here seems to be a bad idea, because
6212 sometimes we need a node as a input and did not want that
6214 However, it might be useful to move this into a later phase
6215 (if you think that optimizing such code is useful). */
6216 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
6217 return new_Unknown(get_irn_mode(node));
6222 /* With this code we violate the agreement that local_optimize
6223 only leaves Bads in Block, Phi and Tuple nodes. */
6224 /* If Block has only Bads as predecessors it's garbage. */
6225 /* If Phi has only Bads as predecessors it's garbage. */
6226 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
6227 irn_arity = get_irn_arity(node);
6228 for (i = 0; i < irn_arity; i++) {
6229 if (!is_Bad(get_irn_n(node, i))) break;
6231 if (i == irn_arity) node = new_Bad();
6238 * These optimizations deallocate nodes from the obstack.
6239 * It can only be called if it is guaranteed that no other nodes
6240 * reference this one, i.e., right after construction of a node.
6242 * @param n The node to optimize
6244 * current_ir_graph must be set to the graph of the node!
6246 ir_node *optimize_node(ir_node *n) {
6249 ir_opcode iro = get_irn_opcode(n);
6251 /* Always optimize Phi nodes: part of the construction. */
6252 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6254 /* constant expression evaluation / constant folding */
6255 if (get_opt_constant_folding()) {
6256 /* neither constants nor Tuple values can be evaluated */
6257 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6258 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6259 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6260 /* try to evaluate */
6261 tv = computed_value(n);
6262 if (tv != tarval_bad) {
6264 ir_type *old_tp = get_irn_type(n);
6265 int i, arity = get_irn_arity(n);
6269 * Try to recover the type of the new expression.
6271 for (i = 0; i < arity && !old_tp; ++i)
6272 old_tp = get_irn_type(get_irn_n(n, i));
6275 * we MUST copy the node here temporary, because it's still needed
6276 * for DBG_OPT_CSTEVAL
6278 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6279 oldn = alloca(node_size);
6281 memcpy(oldn, n, node_size);
6282 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6284 /* ARG, copy the in array, we need it for statistics */
6285 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6287 /* note the inplace edges module */
6288 edges_node_deleted(n, current_ir_graph);
6290 /* evaluation was successful -- replace the node. */
6291 irg_kill_node(current_ir_graph, n);
6292 nw = new_Const(get_tarval_mode(tv), tv);
6294 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6295 set_Const_type(nw, old_tp);
6296 DBG_OPT_CSTEVAL(oldn, nw);
6297 tarval_enable_fp_ops(old_fp_mode);
6300 tarval_enable_fp_ops(old_fp_mode);
6304 /* remove unnecessary nodes */
6305 if (get_opt_algebraic_simplification() ||
6306 (iro == iro_Phi) || /* always optimize these nodes. */
6308 (iro == iro_Proj) ||
6309 (iro == iro_Block) ) /* Flags tested local. */
6310 n = equivalent_node(n);
6312 /* Common Subexpression Elimination.
6314 * Checks whether n is already available.
6315 * The block input is used to distinguish different subexpressions. Right
6316 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6317 * subexpressions within a block.
6320 n = identify_cons(current_ir_graph->value_table, n);
6323 edges_node_deleted(oldn, current_ir_graph);
6325 /* We found an existing, better node, so we can deallocate the old node. */
6326 irg_kill_node(current_ir_graph, oldn);
6330 /* Some more constant expression evaluation that does not allow to
6332 iro = get_irn_opcode(n);
6333 if (get_opt_algebraic_simplification() ||
6334 (iro == iro_Cond) ||
6335 (iro == iro_Proj)) /* Flags tested local. */
6336 n = transform_node(n);
6338 /* Remove nodes with dead (Bad) input.
6339 Run always for transformation induced Bads. */
6342 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6343 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6344 n = identify_remember(current_ir_graph->value_table, n);
6348 } /* optimize_node */
6352 * These optimizations never deallocate nodes (in place). This can cause dead
6353 * nodes lying on the obstack. Remove these by a dead node elimination,
6354 * i.e., a copying garbage collection.
6356 ir_node *optimize_in_place_2(ir_node *n) {
6359 ir_opcode iro = get_irn_opcode(n);
6361 if (!get_opt_optimize() && !is_Phi(n)) return n;
6363 /* constant expression evaluation / constant folding */
6364 if (get_opt_constant_folding()) {
6365 /* neither constants nor Tuple values can be evaluated */
6366 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6367 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6368 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6369 /* try to evaluate */
6370 tv = computed_value(n);
6371 if (tv != tarval_bad) {
6372 /* evaluation was successful -- replace the node. */
6373 ir_type *old_tp = get_irn_type(n);
6374 int i, arity = get_irn_arity(n);
6377 * Try to recover the type of the new expression.
6379 for (i = 0; i < arity && !old_tp; ++i)
6380 old_tp = get_irn_type(get_irn_n(n, i));
6382 n = new_Const(get_tarval_mode(tv), tv);
6384 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6385 set_Const_type(n, old_tp);
6387 DBG_OPT_CSTEVAL(oldn, n);
6388 tarval_enable_fp_ops(old_fp_mode);
6391 tarval_enable_fp_ops(old_fp_mode);
6395 /* remove unnecessary nodes */
6396 if (get_opt_constant_folding() ||
6397 (iro == iro_Phi) || /* always optimize these nodes. */
6398 (iro == iro_Id) || /* ... */
6399 (iro == iro_Proj) || /* ... */
6400 (iro == iro_Block) ) /* Flags tested local. */
6401 n = equivalent_node(n);
6403 /** common subexpression elimination **/
6404 /* Checks whether n is already available. */
6405 /* The block input is used to distinguish different subexpressions. Right
6406 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6407 subexpressions within a block. */
6408 if (get_opt_cse()) {
6409 n = identify_remember(current_ir_graph->value_table, n);
6412 /* Some more constant expression evaluation. */
6413 iro = get_irn_opcode(n);
6414 if (get_opt_constant_folding() ||
6415 (iro == iro_Cond) ||
6416 (iro == iro_Proj)) /* Flags tested local. */
6417 n = transform_node(n);
6419 /* Remove nodes with dead (Bad) input.
6420 Run always for transformation induced Bads. */
6423 /* Now we can verify the node, as it has no dead inputs any more. */
6426 /* Now we have a legal, useful node. Enter it in hash table for cse.
6427 Blocks should be unique anyways. (Except the successor of start:
6428 is cse with the start block!) */
6429 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
6430 n = identify_remember(current_ir_graph->value_table, n);
6433 } /* optimize_in_place_2 */
6436 * Wrapper for external use, set proper status bits after optimization.
6438 ir_node *optimize_in_place(ir_node *n) {
6439 /* Handle graph state */
6440 assert(get_irg_phase_state(current_ir_graph) != phase_building);
6442 if (get_opt_global_cse())
6443 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6444 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
6445 set_irg_outs_inconsistent(current_ir_graph);
6447 /* FIXME: Maybe we could also test whether optimizing the node can
6448 change the control graph. */
6449 set_irg_doms_inconsistent(current_ir_graph);
6450 return optimize_in_place_2(n);
6451 } /* optimize_in_place */
6454 * Calculate a hash value of a Const node.
6456 static unsigned hash_Const(const ir_node *node) {
6459 /* special value for const, as they only differ in their tarval. */
6460 h = HASH_PTR(node->attr.con.tv);
6461 h = 9*h + HASH_PTR(get_irn_mode(node));
6467 * Calculate a hash value of a SymConst node.
6469 static unsigned hash_SymConst(const ir_node *node) {
6472 /* special value for const, as they only differ in their symbol. */
6473 h = HASH_PTR(node->attr.symc.sym.type_p);
6474 h = 9*h + HASH_PTR(get_irn_mode(node));
6477 } /* hash_SymConst */
6480 * Set the default hash operation in an ir_op_ops.
6482 * @param code the opcode for the default operation
6483 * @param ops the operations initialized
6488 static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops)
6492 ops->hash = hash_##a; \
6495 /* hash function already set */
6496 if (ops->hash != NULL)
6503 /* use input/mode default hash if no function was given */
6504 ops->hash = firm_default_hash;
6512 * Sets the default operation for an ir_ops.
6514 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
6515 ops = firm_set_default_hash(code, ops);
6516 ops = firm_set_default_computed_value(code, ops);
6517 ops = firm_set_default_equivalent_node(code, ops);
6518 ops = firm_set_default_transform_node(code, ops);
6519 ops = firm_set_default_node_cmp_attr(code, ops);
6520 ops = firm_set_default_get_type(code, ops);
6521 ops = firm_set_default_get_type_attr(code, ops);
6522 ops = firm_set_default_get_entity_attr(code, ops);
6525 } /* firm_set_default_operations */