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
31 #include "irgraph_t.h"
32 #include "iredges_t.h"
39 #include "dbginfo_t.h"
40 #include "iropt_dbg.h"
45 #include "opt_confirms.h"
46 #include "opt_polymorphy.h"
51 #include "firm_types.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 (! mode_is_float(mode)) {
144 return get_mode_null(mode);
150 if ((ta != tarval_bad) && (tb != tarval_bad))
151 return tarval_sub(ta, tb, mode);
154 } /* computed_value_Sub */
157 * Return the value of a Carry.
158 * Special : a op 0, 0 op b
160 static tarval *computed_value_Carry(const ir_node *n) {
161 ir_node *a = get_binop_left(n);
162 ir_node *b = get_binop_right(n);
163 ir_mode *m = get_irn_mode(n);
165 tarval *ta = value_of(a);
166 tarval *tb = value_of(b);
168 if ((ta != tarval_bad) && (tb != tarval_bad)) {
170 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
172 if (tarval_is_null(ta) || tarval_is_null(tb))
173 return get_mode_null(m);
176 } /* computed_value_Carry */
179 * Return the value of a Borrow.
182 static tarval *computed_value_Borrow(const ir_node *n) {
183 ir_node *a = get_binop_left(n);
184 ir_node *b = get_binop_right(n);
185 ir_mode *m = get_irn_mode(n);
187 tarval *ta = value_of(a);
188 tarval *tb = value_of(b);
190 if ((ta != tarval_bad) && (tb != tarval_bad)) {
191 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
192 } else if (tarval_is_null(ta)) {
193 return get_mode_null(m);
196 } /* computed_value_Borrow */
199 * Return the value of an unary Minus.
201 static tarval *computed_value_Minus(const ir_node *n) {
202 ir_node *a = get_Minus_op(n);
203 tarval *ta = value_of(a);
205 if (ta != tarval_bad)
206 return tarval_neg(ta);
209 } /* computed_value_Minus */
212 * Return the value of a Mul.
214 static tarval *computed_value_Mul(const ir_node *n) {
215 ir_node *a = get_Mul_left(n);
216 ir_node *b = get_Mul_right(n);
219 tarval *ta = value_of(a);
220 tarval *tb = value_of(b);
222 mode = get_irn_mode(n);
223 if (mode != get_irn_mode(a)) {
224 /* n * n = 2n bit multiplication */
225 ta = tarval_convert_to(ta, mode);
226 tb = tarval_convert_to(tb, mode);
229 if (ta != tarval_bad && tb != tarval_bad) {
230 return tarval_mul(ta, tb);
232 /* a * 0 != 0 if a == NaN or a == Inf */
233 if (!mode_is_float(mode)) {
234 /* a*0 = 0 or 0*b = 0 */
235 if (ta == get_mode_null(mode))
237 if (tb == get_mode_null(mode))
242 } /* computed_value_Mul */
245 * Return the value of an Abs.
247 static tarval *computed_value_Abs(const ir_node *n) {
248 ir_node *a = get_Abs_op(n);
249 tarval *ta = value_of(a);
251 if (ta != tarval_bad)
252 return tarval_abs(ta);
255 } /* computed_value_Abs */
258 * Return the value of an And.
259 * Special case: a & 0, 0 & b
261 static tarval *computed_value_And(const ir_node *n) {
262 ir_node *a = get_And_left(n);
263 ir_node *b = get_And_right(n);
265 tarval *ta = value_of(a);
266 tarval *tb = value_of(b);
268 if ((ta != tarval_bad) && (tb != tarval_bad)) {
269 return tarval_and (ta, tb);
271 if (tarval_is_null(ta)) return ta;
272 if (tarval_is_null(tb)) return tb;
275 } /* computed_value_And */
278 * Return the value of an Or.
279 * Special case: a | 1...1, 1...1 | b
281 static tarval *computed_value_Or(const ir_node *n) {
282 ir_node *a = get_Or_left(n);
283 ir_node *b = get_Or_right(n);
285 tarval *ta = value_of(a);
286 tarval *tb = value_of(b);
288 if ((ta != tarval_bad) && (tb != tarval_bad)) {
289 return tarval_or (ta, tb);
291 if (tarval_is_all_one(ta)) return ta;
292 if (tarval_is_all_one(tb)) return tb;
295 } /* computed_value_Or */
298 * Return the value of an Eor.
300 static tarval *computed_value_Eor(const ir_node *n) {
301 ir_node *a = get_Eor_left(n);
302 ir_node *b = get_Eor_right(n);
307 return get_mode_null(get_irn_mode(n));
312 if ((ta != tarval_bad) && (tb != tarval_bad)) {
313 return tarval_eor(ta, tb);
316 } /* computed_value_Eor */
319 * Return the value of a Not.
321 static tarval *computed_value_Not(const ir_node *n) {
322 ir_node *a = get_Not_op(n);
323 tarval *ta = value_of(a);
325 if (ta != tarval_bad)
326 return tarval_not(ta);
329 } /* computed_value_Not */
332 * Return the value of a Shl.
334 static tarval *computed_value_Shl(const ir_node *n) {
335 ir_node *a = get_Shl_left(n);
336 ir_node *b = get_Shl_right(n);
338 tarval *ta = value_of(a);
339 tarval *tb = value_of(b);
341 if ((ta != tarval_bad) && (tb != tarval_bad)) {
342 return tarval_shl(ta, tb);
345 } /* computed_value_Shl */
348 * Return the value of a Shr.
350 static tarval *computed_value_Shr(const ir_node *n) {
351 ir_node *a = get_Shr_left(n);
352 ir_node *b = get_Shr_right(n);
354 tarval *ta = value_of(a);
355 tarval *tb = value_of(b);
357 if ((ta != tarval_bad) && (tb != tarval_bad)) {
358 return tarval_shr(ta, tb);
361 } /* computed_value_Shr */
364 * Return the value of a Shrs.
366 static tarval *computed_value_Shrs(const ir_node *n) {
367 ir_node *a = get_Shrs_left(n);
368 ir_node *b = get_Shrs_right(n);
370 tarval *ta = value_of(a);
371 tarval *tb = value_of(b);
373 if ((ta != tarval_bad) && (tb != tarval_bad)) {
374 return tarval_shrs(ta, tb);
377 } /* computed_value_Shrs */
380 * Return the value of a Rotl.
382 static tarval *computed_value_Rotl(const ir_node *n) {
383 ir_node *a = get_Rotl_left(n);
384 ir_node *b = get_Rotl_right(n);
386 tarval *ta = value_of(a);
387 tarval *tb = value_of(b);
389 if ((ta != tarval_bad) && (tb != tarval_bad)) {
390 return tarval_rotl(ta, tb);
393 } /* computed_value_Rotl */
396 * Return the value of a Conv.
398 static tarval *computed_value_Conv(const ir_node *n) {
399 ir_node *a = get_Conv_op(n);
400 tarval *ta = value_of(a);
402 if (ta != tarval_bad)
403 return tarval_convert_to(ta, get_irn_mode(n));
406 } /* computed_value_Conv */
409 * Calculate the value of a Mux: can be evaluated, if the
410 * sel and the right input are known.
412 static tarval *computed_value_Mux(const ir_node *n) {
413 ir_node *sel = get_Mux_sel(n);
414 tarval *ts = value_of(sel);
416 if (ts == get_tarval_b_true()) {
417 ir_node *v = get_Mux_true(n);
420 else if (ts == get_tarval_b_false()) {
421 ir_node *v = get_Mux_false(n);
425 } /* computed_value_Mux */
428 * Calculate the value of a Confirm: can be evaluated,
429 * if it has the form Confirm(x, '=', Const).
431 static tarval *computed_value_Confirm(const ir_node *n) {
433 * Beware: we might produce Phi(Confirm(x == true), Confirm(x == false)).
434 * Do NOT optimize them away (jump threading wants them), so wait until
435 * remove_confirm is activated.
437 if (get_opt_remove_confirm()) {
438 if (get_Confirm_cmp(n) == pn_Cmp_Eq) {
439 tarval *tv = value_of(get_Confirm_bound(n));
440 if (tv != tarval_bad)
444 return value_of(get_Confirm_value(n));
445 } /* computed_value_Confirm */
448 * Return the value of a Proj(Cmp).
450 * This performs a first step of unreachable code elimination.
451 * Proj can not be computed, but folding a Cmp above the Proj here is
452 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
454 * There are several case where we can evaluate a Cmp node, see later.
456 static tarval *computed_value_Proj_Cmp(const ir_node *n) {
457 ir_node *a = get_Proj_pred(n);
458 ir_node *aa = get_Cmp_left(a);
459 ir_node *ab = get_Cmp_right(a);
460 long proj_nr = get_Proj_proj(n);
463 * BEWARE: a == a is NOT always True for floating Point values, as
464 * NaN != NaN is defined, so we must check this here.
467 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
470 /* This is a trick with the bits used for encoding the Cmp
471 Proj numbers, the following statement is not the same:
472 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
473 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
475 tarval *taa = value_of(aa);
476 tarval *tab = value_of(ab);
477 ir_mode *mode = get_irn_mode(aa);
480 * The predecessors of Cmp are target values. We can evaluate
483 if ((taa != tarval_bad) && (tab != tarval_bad)) {
484 /* strange checks... */
485 pn_Cmp flags = tarval_cmp(taa, tab);
486 if (flags != pn_Cmp_False) {
487 return new_tarval_from_long (proj_nr & flags, mode_b);
490 /* for integer values, we can check against MIN/MAX */
491 else if (mode_is_int(mode)) {
492 /* MIN <=/> x. This results in true/false. */
493 if (taa == get_mode_min(mode)) {
494 /* a compare with the MIN value */
495 if (proj_nr == pn_Cmp_Le)
496 return get_tarval_b_true();
497 else if (proj_nr == pn_Cmp_Gt)
498 return get_tarval_b_false();
500 /* x >=/< MIN. This results in true/false. */
502 if (tab == get_mode_min(mode)) {
503 /* a compare with the MIN value */
504 if (proj_nr == pn_Cmp_Ge)
505 return get_tarval_b_true();
506 else if (proj_nr == pn_Cmp_Lt)
507 return get_tarval_b_false();
509 /* MAX >=/< x. This results in true/false. */
510 else if (taa == get_mode_max(mode)) {
511 if (proj_nr == pn_Cmp_Ge)
512 return get_tarval_b_true();
513 else if (proj_nr == pn_Cmp_Lt)
514 return get_tarval_b_false();
516 /* x <=/> MAX. This results in true/false. */
517 else if (tab == get_mode_max(mode)) {
518 if (proj_nr == pn_Cmp_Le)
519 return get_tarval_b_true();
520 else if (proj_nr == pn_Cmp_Gt)
521 return get_tarval_b_false();
524 pn_Cmp cmp_result = vrp_cmp(aa, ab);
525 if (cmp_result != pn_Cmp_False) {
526 return new_tarval_from_long(cmp_result & proj_nr, mode_b);
530 * The predecessors are Allocs or (void*)(0) constants. Allocs never
531 * return NULL, they raise an exception. Therefore we can predict
535 ir_node *aaa = skip_Proj(aa);
536 ir_node *aba = skip_Proj(ab);
538 if ( ( (/* aa is ProjP and aaa is Alloc */
540 && mode_is_reference(get_irn_mode(aa))
542 && ( (/* ab is NULL */
543 mode_is_reference(get_irn_mode(ab))
544 && tarval_is_null(tab))
545 || (/* ab is other Alloc */
547 && mode_is_reference(get_irn_mode(ab))
550 || (/* aa is NULL and aba is Alloc */
551 mode_is_reference(get_irn_mode(aa))
552 && tarval_is_null(taa)
554 && mode_is_reference(get_irn_mode(ab))
557 return new_tarval_from_long(proj_nr & pn_Cmp_Lg, mode_b);
559 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
560 } /* computed_value_Proj_Cmp */
563 * Return the value of a floating point Quot.
565 static tarval *do_computed_value_Quot(const ir_node *a, const ir_node *b) {
566 tarval *ta = value_of(a);
567 tarval *tb = value_of(b);
569 /* cannot optimize 0 / b = 0 because of NaN */
570 if (ta != tarval_bad && tb != tarval_bad)
571 return tarval_quo(ta, tb);
573 } /* do_computed_value_Quot */
576 * Calculate the value of an integer Div of two nodes.
577 * Special case: 0 / b
579 static tarval *do_computed_value_Div(const ir_node *a, const ir_node *b) {
580 tarval *ta = value_of(a);
582 const ir_node *dummy;
584 /* Compute c1 / c2 or 0 / a, a != 0 */
585 if (tarval_is_null(ta) && value_not_zero(b, &dummy))
586 return ta; /* 0 / b == 0 */
588 if (ta != tarval_bad && tb != tarval_bad)
589 return tarval_div(ta, tb);
591 } /* do_computed_value_Div */
594 * Calculate the value of an integer Mod of two nodes.
595 * Special case: a % 1
597 static tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b) {
598 tarval *ta = value_of(a);
599 tarval *tb = value_of(b);
601 /* Compute a % 1 or c1 % c2 */
602 if (tarval_is_one(tb))
603 return get_mode_null(get_irn_mode(a));
604 if (ta != tarval_bad && tb != tarval_bad)
605 return tarval_mod(ta, tb);
607 } /* do_computed_value_Mod */
610 * Return the value of a Proj(DivMod).
612 static tarval *computed_value_Proj_DivMod(const ir_node *n) {
613 long proj_nr = get_Proj_proj(n);
615 /* compute either the Div or the Mod part */
616 if (proj_nr == pn_DivMod_res_div) {
617 const ir_node *a = get_Proj_pred(n);
618 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
619 } else if (proj_nr == pn_DivMod_res_mod) {
620 const ir_node *a = get_Proj_pred(n);
621 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
624 } /* computed_value_Proj_DivMod */
627 * Return the value of a Proj(Div).
629 static tarval *computed_value_Proj_Div(const ir_node *n) {
630 long proj_nr = get_Proj_proj(n);
632 if (proj_nr == pn_Div_res) {
633 const ir_node *a = get_Proj_pred(n);
634 return do_computed_value_Div(get_Div_left(a), get_Div_right(a));
637 } /* computed_value_Proj_Div */
640 * Return the value of a Proj(Mod).
642 static tarval *computed_value_Proj_Mod(const ir_node *n) {
643 long proj_nr = get_Proj_proj(n);
645 if (proj_nr == pn_Mod_res) {
646 const ir_node *a = get_Proj_pred(n);
647 return do_computed_value_Mod(get_Mod_left(a), get_Mod_right(a));
650 } /* computed_value_Proj_Mod */
653 * Return the value of a Proj(Quot).
655 static tarval *computed_value_Proj_Quot(const ir_node *n) {
656 long proj_nr = get_Proj_proj(n);
658 if (proj_nr == pn_Quot_res) {
659 const ir_node *a = get_Proj_pred(n);
660 return do_computed_value_Quot(get_Quot_left(a), get_Quot_right(a));
663 } /* computed_value_Proj_Quot */
666 * Return the value of a Proj.
668 static tarval *computed_value_Proj(const ir_node *proj) {
669 ir_node *n = get_Proj_pred(proj);
671 if (n->op->ops.computed_value_Proj != NULL)
672 return n->op->ops.computed_value_Proj(proj);
674 } /* computed_value_Proj */
677 * If the parameter n can be computed, return its value, else tarval_bad.
678 * Performs constant folding.
680 * @param n The node this should be evaluated
682 tarval *computed_value(const ir_node *n) {
683 if(mode_is_int(get_irn_mode(n)) && tarval_is_all_one(
684 tarval_or(n->vrp.bits_set, n->vrp.bits_not_set))) {
685 return n->vrp.bits_set;
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; \
746 } /* firm_set_default_computed_value */
749 * Returns a equivalent block for another block.
750 * If the block has only one predecessor, this is
751 * the equivalent one. If the only predecessor of a block is
752 * the block itself, this is a dead block.
754 * If both predecessors of a block are the branches of a binary
755 * Cond, the equivalent block is Cond's block.
757 * If all predecessors of a block are bad or lies in a dead
758 * block, the current block is dead as well.
760 * Note, that blocks are NEVER turned into Bad's, instead
761 * the dead_block flag is set. So, never test for is_Bad(block),
762 * always use is_dead_Block(block).
764 static ir_node *equivalent_node_Block(ir_node *n)
769 /* don't optimize dead or labeled blocks */
770 if (is_Block_dead(n) || has_Block_entity(n))
773 n_preds = get_Block_n_cfgpreds(n);
775 /* The Block constructor does not call optimize, but mature_immBlock()
776 calls the optimization. */
777 assert(get_Block_matured(n));
779 /* Straightening: a single entry Block following a single exit Block
780 can be merged, if it is not the Start block. */
781 /* !!! Beware, all Phi-nodes of n must have been optimized away.
782 This should be true, as the block is matured before optimize is called.
783 But what about Phi-cycles with the Phi0/Id that could not be resolved?
784 Remaining Phi nodes are just Ids. */
786 ir_node *pred = skip_Proj(get_Block_cfgpred(n, 0));
789 ir_node *predblock = get_nodes_block(pred);
790 if (predblock == oldn) {
791 /* Jmp jumps into the block it is in -- deal self cycle. */
792 n = set_Block_dead(n);
793 DBG_OPT_DEAD_BLOCK(oldn, n);
794 } else if (get_opt_control_flow_straightening()) {
796 DBG_OPT_STG(oldn, n);
798 } else if (is_Cond(pred)) {
799 ir_node *predblock = get_nodes_block(pred);
800 if (predblock == oldn) {
801 /* Jmp jumps into the block it is in -- deal self cycle. */
802 n = set_Block_dead(n);
803 DBG_OPT_DEAD_BLOCK(oldn, n);
806 } else if ((n_preds == 2) &&
807 (get_opt_control_flow_weak_simplification())) {
808 /* Test whether Cond jumps twice to this block
809 * The more general case which more than 2 predecessors is handles
810 * in optimize_cf(), we handle only this special case for speed here.
812 ir_node *a = get_Block_cfgpred(n, 0);
813 ir_node *b = get_Block_cfgpred(n, 1);
815 if (is_Proj(a) && is_Proj(b)) {
816 ir_node *cond = get_Proj_pred(a);
818 if (cond == get_Proj_pred(b) && is_Cond(cond) &&
819 get_irn_mode(get_Cond_selector(cond)) == mode_b) {
820 /* Also a single entry Block following a single exit Block. Phis have
821 twice the same operand and will be optimized away. */
822 n = get_nodes_block(cond);
823 DBG_OPT_IFSIM1(oldn, a, b, n);
826 } else if (get_opt_unreachable_code() &&
827 (n != get_irg_start_block(current_ir_graph)) &&
828 (n != get_irg_end_block(current_ir_graph)) ) {
831 /* If all inputs are dead, this block is dead too, except if it is
832 the start or end block. This is one step of unreachable code
834 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
835 ir_node *pred = get_Block_cfgpred(n, i);
838 if (is_Bad(pred)) continue;
839 pred_blk = get_nodes_block(skip_Proj(pred));
841 if (is_Block_dead(pred_blk)) continue;
844 /* really found a living input */
849 n = set_Block_dead(n);
850 DBG_OPT_DEAD_BLOCK(oldn, n);
855 } /* equivalent_node_Block */
858 * Returns a equivalent node for a Jmp, a Bad :-)
859 * Of course this only happens if the Block of the Jmp is dead.
861 static ir_node *equivalent_node_Jmp(ir_node *n) {
864 /* unreachable code elimination */
865 if (is_Block_dead(get_nodes_block(n))) {
866 n = get_irg_bad(current_ir_graph);
867 DBG_OPT_DEAD_BLOCK(oldn, n);
870 } /* equivalent_node_Jmp */
872 /** Raise is handled in the same way as Jmp. */
873 #define equivalent_node_Raise equivalent_node_Jmp
876 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
877 See transform_node_Proj_Cond(). */
880 * Optimize operations that are commutative and have neutral 0,
881 * so a op 0 = 0 op a = a.
883 static ir_node *equivalent_node_neutral_zero(ir_node *n) {
886 ir_node *a = get_binop_left(n);
887 ir_node *b = get_binop_right(n);
892 /* After running compute_node there is only one constant predecessor.
893 Find this predecessors value and remember the other node: */
894 if ((tv = value_of(a)) != tarval_bad) {
896 } else if ((tv = value_of(b)) != tarval_bad) {
901 /* If this predecessors constant value is zero, the operation is
902 * unnecessary. Remove it.
904 * Beware: If n is a Add, the mode of on and n might be different
905 * which happens in this rare construction: NULL + 3.
906 * Then, a Conv would be needed which we cannot include here.
908 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
911 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
915 } /* equivalent_node_neutral_zero */
918 * Eor is commutative and has neutral 0.
920 static ir_node *equivalent_node_Eor(ir_node *n) {
925 n = equivalent_node_neutral_zero(n);
926 if (n != oldn) return n;
929 b = get_Eor_right(n);
932 ir_node *aa = get_Eor_left(a);
933 ir_node *ab = get_Eor_right(a);
936 /* (a ^ b) ^ a -> b */
938 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
940 } else if (ab == b) {
941 /* (a ^ b) ^ b -> a */
943 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
948 ir_node *ba = get_Eor_left(b);
949 ir_node *bb = get_Eor_right(b);
952 /* a ^ (a ^ b) -> b */
954 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
956 } else if (bb == a) {
957 /* a ^ (b ^ a) -> b */
959 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
967 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
969 * The second one looks strange, but this construct
970 * is used heavily in the LCC sources :-).
972 * Beware: The Mode of an Add may be different than the mode of its
973 * predecessors, so we could not return a predecessors in all cases.
975 static ir_node *equivalent_node_Add(ir_node *n) {
977 ir_node *left, *right;
978 ir_mode *mode = get_irn_mode(n);
980 n = equivalent_node_neutral_zero(n);
984 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
985 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
988 left = get_Add_left(n);
989 right = get_Add_right(n);
992 if (get_Sub_right(left) == right) {
995 n = get_Sub_left(left);
996 if (mode == get_irn_mode(n)) {
997 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
1002 if (is_Sub(right)) {
1003 if (get_Sub_right(right) == left) {
1006 n = get_Sub_left(right);
1007 if (mode == get_irn_mode(n)) {
1008 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
1014 } /* equivalent_node_Add */
1017 * optimize operations that are not commutative but have neutral 0 on left,
1020 static ir_node *equivalent_node_left_zero(ir_node *n) {
1023 ir_node *a = get_binop_left(n);
1024 ir_node *b = get_binop_right(n);
1025 tarval *tb = value_of(b);
1027 if (tarval_is_null(tb)) {
1030 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1033 } /* equivalent_node_left_zero */
1035 #define equivalent_node_Shl equivalent_node_left_zero
1036 #define equivalent_node_Shr equivalent_node_left_zero
1037 #define equivalent_node_Shrs equivalent_node_left_zero
1038 #define equivalent_node_Rotl equivalent_node_left_zero
1041 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
1043 * The second one looks strange, but this construct
1044 * is used heavily in the LCC sources :-).
1046 * Beware: The Mode of a Sub may be different than the mode of its
1047 * predecessors, so we could not return a predecessors in all cases.
1049 static ir_node *equivalent_node_Sub(ir_node *n) {
1052 ir_mode *mode = get_irn_mode(n);
1055 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1056 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1059 b = get_Sub_right(n);
1062 /* Beware: modes might be different */
1063 if (tarval_is_null(tb)) {
1064 ir_node *a = get_Sub_left(n);
1065 if (mode == get_irn_mode(a)) {
1068 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1072 } /* equivalent_node_Sub */
1076 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
1079 * -(-a) == a, but might overflow two times.
1080 * We handle it anyway here but the better way would be a
1081 * flag. This would be needed for Pascal for instance.
1083 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
1085 ir_node *pred = get_unop_op(n);
1087 /* optimize symmetric unop */
1088 if (get_irn_op(pred) == get_irn_op(n)) {
1089 n = get_unop_op(pred);
1090 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
1093 } /* equivalent_node_idempotent_unop */
1095 /** Optimize Not(Not(x)) == x. */
1096 #define equivalent_node_Not equivalent_node_idempotent_unop
1098 /** -(-x) == x ??? Is this possible or can --x raise an
1099 out of bounds exception if min =! max? */
1100 #define equivalent_node_Minus equivalent_node_idempotent_unop
1103 * Optimize a * 1 = 1 * a = a.
1105 static ir_node *equivalent_node_Mul(ir_node *n) {
1107 ir_node *a = get_Mul_left(n);
1109 /* we can handle here only the n * n = n bit cases */
1110 if (get_irn_mode(n) == get_irn_mode(a)) {
1111 ir_node *b = get_Mul_right(n);
1115 * Mul is commutative and has again an other neutral element.
1116 * Constants are place right, so check this case first.
1119 if (tarval_is_one(tv)) {
1121 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1124 if (tarval_is_one(tv)) {
1126 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1131 } /* equivalent_node_Mul */
1134 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1136 static ir_node *equivalent_node_Or(ir_node *n) {
1139 ir_node *a = get_Or_left(n);
1140 ir_node *b = get_Or_right(n);
1144 n = a; /* Or has it's own neutral element */
1145 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1148 /* constants are cormalized to right, check this site first */
1150 if (tarval_is_null(tv)) {
1152 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1156 if (tarval_is_null(tv)) {
1158 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1163 } /* equivalent_node_Or */
1166 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1168 static ir_node *equivalent_node_And(ir_node *n) {
1171 ir_node *a = get_And_left(n);
1172 ir_node *b = get_And_right(n);
1176 n = a; /* And has it's own neutral element */
1177 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1180 /* constants are normalized to right, check this site first */
1182 if (tarval_is_all_one(tv)) {
1184 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1187 if (tv != get_tarval_bad()) {
1188 ir_mode *mode = get_irn_mode(n);
1189 if (!mode_is_signed(mode) && is_Conv(a)) {
1190 ir_node *convop = get_Conv_op(a);
1191 ir_mode *convopmode = get_irn_mode(convop);
1192 if (!mode_is_signed(convopmode)) {
1193 if (tarval_is_all_one(tarval_convert_to(tv, convopmode))) {
1194 /* Conv(X) & all_one(mode(X)) = Conv(X) */
1196 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1203 if (tarval_is_all_one(tv)) {
1205 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1209 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1212 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1217 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1220 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1225 } /* equivalent_node_And */
1228 * Try to remove useless Conv's:
1230 static ir_node *equivalent_node_Conv(ir_node *n) {
1232 ir_node *a = get_Conv_op(n);
1234 ir_mode *n_mode = get_irn_mode(n);
1235 ir_mode *a_mode = get_irn_mode(a);
1238 if (n_mode == a_mode) { /* No Conv necessary */
1239 if (get_Conv_strict(n)) {
1242 /* neither Minus nor Abs nor Confirm change the precision,
1243 so we can "look-through" */
1246 p = get_Minus_op(p);
1247 } else if (is_Abs(p)) {
1249 } else if (is_Confirm(p)) {
1250 p = get_Confirm_value(p);
1256 if (is_Conv(p) && get_Conv_strict(p)) {
1257 /* we known already, that a_mode == n_mode, and neither
1258 Abs nor Minus change the mode, so the second Conv
1260 assert(get_irn_mode(p) == n_mode);
1262 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1266 ir_node *pred = get_Proj_pred(p);
1267 if (is_Load(pred)) {
1268 /* Loads always return with the exact precision of n_mode */
1269 assert(get_Load_mode(pred) == n_mode);
1271 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1274 if (is_Proj(pred) && get_Proj_proj(pred) == pn_Start_T_args) {
1275 pred = get_Proj_pred(pred);
1276 if (is_Start(pred)) {
1277 /* Arguments always return with the exact precision,
1278 as strictConv's are place before Call -- if the
1279 caller was compiled with the same setting.
1280 Otherwise, the semantics is probably still right. */
1281 assert(get_irn_mode(p) == n_mode);
1283 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1289 /* special case: the immediate predecessor is also a Conv */
1290 if (! get_Conv_strict(a)) {
1291 /* first one is not strict, kick it */
1293 a_mode = get_irn_mode(a);
1297 /* else both are strict conv, second is superfluous */
1299 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1304 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1307 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1308 ir_node *b = get_Conv_op(a);
1309 ir_mode *b_mode = get_irn_mode(b);
1311 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1312 /* both are strict conv */
1313 if (smaller_mode(a_mode, n_mode)) {
1314 /* both are strict, but the first is smaller, so
1315 the second cannot remove more precision, remove the
1317 set_Conv_strict(n, 0);
1320 if (n_mode == b_mode) {
1321 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1322 if (n_mode == mode_b) {
1323 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1324 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1326 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1327 if (values_in_mode(b_mode, a_mode)) {
1328 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1329 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1334 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1335 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1336 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1337 unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode);
1339 if (float_mantissa >= int_mantissa) {
1341 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1346 if (smaller_mode(b_mode, a_mode)) {
1347 if (get_Conv_strict(n))
1348 set_Conv_strict(b, 1);
1349 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1350 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1357 } /* equivalent_node_Conv */
1360 * A Cast may be removed if the type of the previous node
1361 * is already the type of the Cast.
1363 static ir_node *equivalent_node_Cast(ir_node *n) {
1365 ir_node *pred = get_Cast_op(n);
1367 if (get_irn_type(pred) == get_Cast_type(n)) {
1369 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1372 } /* equivalent_node_Cast */
1375 * - fold Phi-nodes, iff they have only one predecessor except
1378 static ir_node *equivalent_node_Phi(ir_node *n) {
1383 ir_node *first_val = NULL; /* to shutup gcc */
1385 if (!get_opt_normalize()) return n;
1387 n_preds = get_Phi_n_preds(n);
1389 block = get_nodes_block(n);
1390 if (is_Block_dead(block)) /* Control dead */
1391 return get_irg_bad(current_ir_graph);
1393 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1395 /* Find first non-self-referencing input */
1396 for (i = 0; i < n_preds; ++i) {
1397 first_val = get_Phi_pred(n, i);
1398 if ( (first_val != n) /* not self pointer */
1400 /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad
1401 * predecessors. Then, Phi nodes in dead code might be removed, causing
1402 * nodes pointing to themself (Add's for instance).
1403 * This is really bad and causes endless recursions in several
1404 * code pathes, so we do NOT optimize such a code.
1405 * This is not that bad as it sounds, optimize_cf() removes bad control flow
1406 * (and bad Phi predecessors), so live code is optimized later.
1408 && (! is_Bad(get_Block_cfgpred(block, i)))
1410 ) { /* value not dead */
1411 break; /* then found first value. */
1416 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1417 return get_irg_bad(current_ir_graph);
1420 /* search for rest of inputs, determine if any of these
1421 are non-self-referencing */
1422 while (++i < n_preds) {
1423 ir_node *scnd_val = get_Phi_pred(n, i);
1424 if ( (scnd_val != n)
1425 && (scnd_val != first_val)
1428 && (! is_Bad(get_Block_cfgpred(block, i)))
1436 /* Fold, if no multiple distinct non-self-referencing inputs */
1438 DBG_OPT_PHI(oldn, n);
1441 } /* equivalent_node_Phi */
1444 * Several optimizations:
1445 * - fold Sync-nodes, iff they have only one predecessor except
1448 static ir_node *equivalent_node_Sync(ir_node *n) {
1449 int arity = get_Sync_n_preds(n);
1452 for (i = 0; i < arity;) {
1453 ir_node *pred = get_Sync_pred(n, i);
1456 /* Remove Bad predecessors */
1463 /* Remove duplicate predecessors */
1469 if (get_Sync_pred(n, j) == pred) {
1477 if (arity == 0) return get_irg_bad(current_ir_graph);
1478 if (arity == 1) return get_Sync_pred(n, 0);
1480 } /* equivalent_node_Sync */
1483 * Optimize Proj(Tuple).
1485 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj) {
1486 ir_node *oldn = proj;
1487 ir_node *tuple = get_Proj_pred(proj);
1489 /* Remove the Tuple/Proj combination. */
1490 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1491 DBG_OPT_TUPLE(oldn, tuple, proj);
1494 } /* equivalent_node_Proj_Tuple */
1497 * Optimize a / 1 = a.
1499 static ir_node *equivalent_node_Proj_Div(ir_node *proj) {
1500 ir_node *oldn = proj;
1501 ir_node *div = get_Proj_pred(proj);
1502 ir_node *b = get_Div_right(div);
1503 tarval *tb = value_of(b);
1505 /* Div is not commutative. */
1506 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1507 switch (get_Proj_proj(proj)) {
1509 proj = get_Div_mem(div);
1510 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1514 proj = get_Div_left(div);
1515 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1519 /* we cannot replace the exception Proj's here, this is done in
1520 transform_node_Proj_Div() */
1525 } /* equivalent_node_Proj_Div */
1528 * Optimize a / 1.0 = a.
1530 static ir_node *equivalent_node_Proj_Quot(ir_node *proj) {
1531 ir_node *oldn = proj;
1532 ir_node *quot = get_Proj_pred(proj);
1533 ir_node *b = get_Quot_right(quot);
1534 tarval *tb = value_of(b);
1536 /* Div is not commutative. */
1537 if (tarval_is_one(tb)) { /* Quot(x, 1) == x */
1538 switch (get_Proj_proj(proj)) {
1540 proj = get_Quot_mem(quot);
1541 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1545 proj = get_Quot_left(quot);
1546 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1550 /* we cannot replace the exception Proj's here, this is done in
1551 transform_node_Proj_Quot() */
1556 } /* equivalent_node_Proj_Quot */
1559 * Optimize a / 1 = a.
1561 static ir_node *equivalent_node_Proj_DivMod(ir_node *proj) {
1562 ir_node *oldn = proj;
1563 ir_node *divmod = get_Proj_pred(proj);
1564 ir_node *b = get_DivMod_right(divmod);
1565 tarval *tb = value_of(b);
1567 /* Div is not commutative. */
1568 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1569 switch (get_Proj_proj(proj)) {
1571 proj = get_DivMod_mem(divmod);
1572 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1575 case pn_DivMod_res_div:
1576 proj = get_DivMod_left(divmod);
1577 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1581 /* we cannot replace the exception Proj's here, this is done in
1582 transform_node_Proj_DivMod().
1583 Note further that the pn_DivMod_res_div case is handled in
1584 computed_value_Proj(). */
1589 } /* equivalent_node_Proj_DivMod */
1592 * Optimize CopyB(mem, x, x) into a Nop.
1594 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj) {
1595 ir_node *oldn = proj;
1596 ir_node *copyb = get_Proj_pred(proj);
1597 ir_node *a = get_CopyB_dst(copyb);
1598 ir_node *b = get_CopyB_src(copyb);
1601 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1602 switch (get_Proj_proj(proj)) {
1603 case pn_CopyB_M_regular:
1604 proj = get_CopyB_mem(copyb);
1605 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1608 case pn_CopyB_X_except:
1609 DBG_OPT_EXC_REM(proj);
1610 proj = get_irg_bad(current_ir_graph);
1615 } /* equivalent_node_Proj_CopyB */
1618 * Optimize Bounds(idx, idx, upper) into idx.
1620 static ir_node *equivalent_node_Proj_Bound(ir_node *proj) {
1621 ir_node *oldn = proj;
1622 ir_node *bound = get_Proj_pred(proj);
1623 ir_node *idx = get_Bound_index(bound);
1624 ir_node *pred = skip_Proj(idx);
1627 if (idx == get_Bound_lower(bound))
1629 else if (is_Bound(pred)) {
1631 * idx was Bounds checked in the same MacroBlock previously,
1632 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1634 ir_node *lower = get_Bound_lower(bound);
1635 ir_node *upper = get_Bound_upper(bound);
1636 if (get_Bound_lower(pred) == lower &&
1637 get_Bound_upper(pred) == upper &&
1638 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
1640 * One could expect that we simply return the previous
1641 * Bound here. However, this would be wrong, as we could
1642 * add an exception Proj to a new location then.
1643 * So, we must turn in into a tuple.
1649 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1650 switch (get_Proj_proj(proj)) {
1652 DBG_OPT_EXC_REM(proj);
1653 proj = get_Bound_mem(bound);
1655 case pn_Bound_X_except:
1656 DBG_OPT_EXC_REM(proj);
1657 proj = get_irg_bad(current_ir_graph);
1661 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1664 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1669 } /* equivalent_node_Proj_Bound */
1672 * Optimize an Exception Proj(Load) with a non-null address.
1674 static ir_node *equivalent_node_Proj_Load(ir_node *proj) {
1675 if (get_opt_ldst_only_null_ptr_exceptions()) {
1676 if (get_irn_mode(proj) == mode_X) {
1677 ir_node *load = get_Proj_pred(proj);
1679 /* get the Load address */
1680 const ir_node *addr = get_Load_ptr(load);
1681 const ir_node *confirm;
1683 if (value_not_null(addr, &confirm)) {
1684 if (get_Proj_proj(proj) == pn_Load_X_except) {
1685 DBG_OPT_EXC_REM(proj);
1686 return get_irg_bad(current_ir_graph);
1692 } /* equivalent_node_Proj_Load */
1695 * Optimize an Exception Proj(Store) with a non-null address.
1697 static ir_node *equivalent_node_Proj_Store(ir_node *proj) {
1698 if (get_opt_ldst_only_null_ptr_exceptions()) {
1699 if (get_irn_mode(proj) == mode_X) {
1700 ir_node *store = get_Proj_pred(proj);
1702 /* get the load/store address */
1703 const ir_node *addr = get_Store_ptr(store);
1704 const ir_node *confirm;
1706 if (value_not_null(addr, &confirm)) {
1707 if (get_Proj_proj(proj) == pn_Store_X_except) {
1708 DBG_OPT_EXC_REM(proj);
1709 return get_irg_bad(current_ir_graph);
1715 } /* equivalent_node_Proj_Store */
1718 * Does all optimizations on nodes that must be done on it's Proj's
1719 * because of creating new nodes.
1721 static ir_node *equivalent_node_Proj(ir_node *proj) {
1722 ir_node *n = get_Proj_pred(proj);
1724 if (get_irn_mode(proj) == mode_X) {
1725 if (is_Block_dead(get_nodes_block(n))) {
1726 /* Remove dead control flow -- early gigo(). */
1727 return get_irg_bad(current_ir_graph);
1730 if (n->op->ops.equivalent_node_Proj)
1731 return n->op->ops.equivalent_node_Proj(proj);
1733 } /* equivalent_node_Proj */
1738 static ir_node *equivalent_node_Id(ir_node *n) {
1745 DBG_OPT_ID(oldn, n);
1747 } /* equivalent_node_Id */
1752 static ir_node *equivalent_node_Mux(ir_node *n)
1754 ir_node *oldn = n, *sel = get_Mux_sel(n);
1756 tarval *ts = value_of(sel);
1758 /* Mux(true, f, t) == t */
1759 if (ts == tarval_b_true) {
1760 n = get_Mux_true(n);
1761 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1764 /* Mux(false, f, t) == f */
1765 if (ts == tarval_b_false) {
1766 n = get_Mux_false(n);
1767 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1770 n_t = get_Mux_true(n);
1771 n_f = get_Mux_false(n);
1773 /* Mux(v, x, T) == x */
1774 if (is_Unknown(n_f)) {
1776 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1779 /* Mux(v, T, x) == x */
1780 if (is_Unknown(n_t)) {
1782 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1786 /* Mux(v, x, x) == x */
1789 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1792 if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1793 ir_node *cmp = get_Proj_pred(sel);
1794 long proj_nr = get_Proj_proj(sel);
1795 ir_node *f = get_Mux_false(n);
1796 ir_node *t = get_Mux_true(n);
1799 * Note further that these optimization work even for floating point
1800 * with NaN's because -NaN == NaN.
1801 * However, if +0 and -0 is handled differently, we cannot use the first one.
1804 ir_node *const cmp_l = get_Cmp_left(cmp);
1805 ir_node *const cmp_r = get_Cmp_right(cmp);
1809 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1810 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1812 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1819 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1820 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1822 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1829 * Note: normalization puts the constant on the right side,
1830 * so we check only one case.
1832 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1833 /* Mux(t CMP 0, X, t) */
1834 if (is_Minus(f) && get_Minus_op(f) == t) {
1835 /* Mux(t CMP 0, -t, t) */
1836 if (proj_nr == pn_Cmp_Eq) {
1837 /* Mux(t == 0, -t, t) ==> -t */
1839 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1840 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1841 /* Mux(t != 0, -t, t) ==> t */
1843 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1850 } /* equivalent_node_Mux */
1853 * Remove Confirm nodes if setting is on.
1854 * Replace Confirms(x, '=', Constlike) by Constlike.
1856 static ir_node *equivalent_node_Confirm(ir_node *n) {
1857 ir_node *pred = get_Confirm_value(n);
1858 pn_Cmp pnc = get_Confirm_cmp(n);
1860 while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1862 * rare case: two identical Confirms one after another,
1863 * replace the second one with the first.
1866 pred = get_Confirm_value(n);
1867 pnc = get_Confirm_cmp(n);
1869 if (get_opt_remove_confirm())
1870 return get_Confirm_value(n);
1875 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1876 * perform no actual computation, as, e.g., the Id nodes. It does not create
1877 * new nodes. It is therefore safe to free n if the node returned is not n.
1878 * If a node returns a Tuple we can not just skip it. If the size of the
1879 * in array fits, we transform n into a tuple (e.g., Div).
1881 ir_node *equivalent_node(ir_node *n) {
1882 if (n->op->ops.equivalent_node)
1883 return n->op->ops.equivalent_node(n);
1885 } /* equivalent_node */
1888 * Sets the default equivalent node operation for an ir_op_ops.
1890 * @param code the opcode for the default operation
1891 * @param ops the operations initialized
1896 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1900 ops->equivalent_node = equivalent_node_##a; \
1902 #define CASE_PROJ(a) \
1904 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1946 } /* firm_set_default_equivalent_node */
1949 * Returns non-zero if a node is a Phi node
1950 * with all predecessors constant.
1952 static int is_const_Phi(ir_node *n) {
1955 if (! is_Phi(n) || get_irn_arity(n) == 0)
1957 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1958 if (! is_Const(get_irn_n(n, i)))
1962 } /* is_const_Phi */
1964 typedef tarval *(*tarval_sub_type)(tarval *a, tarval *b, ir_mode *mode);
1965 typedef tarval *(*tarval_binop_type)(tarval *a, tarval *b);
1968 * in reality eval_func should be tarval (*eval_func)() but incomplete
1969 * declarations are bad style and generate noisy warnings
1971 typedef void (*eval_func)(void);
1974 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1976 static tarval *do_eval(eval_func eval, tarval *a, tarval *b, ir_mode *mode)
1978 if (eval == (eval_func) tarval_sub) {
1979 tarval_sub_type func = (tarval_sub_type)eval;
1981 return func(a, b, mode);
1983 tarval_binop_type func = (tarval_binop_type)eval;
1990 * Apply an evaluator on a binop with a constant operators (and one Phi).
1992 * @param phi the Phi node
1993 * @param other the other operand
1994 * @param eval an evaluator function
1995 * @param mode the mode of the result, may be different from the mode of the Phi!
1996 * @param left if non-zero, other is the left operand, else the right
1998 * @return a new Phi node if the conversion was successful, NULL else
2000 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, eval_func eval, ir_mode *mode, int left) {
2005 int i, n = get_irn_arity(phi);
2007 NEW_ARR_A(void *, res, n);
2009 for (i = 0; i < n; ++i) {
2010 pred = get_irn_n(phi, i);
2011 tv = get_Const_tarval(pred);
2012 tv = do_eval(eval, other, tv, mode);
2014 if (tv == tarval_bad) {
2015 /* folding failed, bad */
2021 for (i = 0; i < n; ++i) {
2022 pred = get_irn_n(phi, i);
2023 tv = get_Const_tarval(pred);
2024 tv = do_eval(eval, tv, other, mode);
2026 if (tv == tarval_bad) {
2027 /* folding failed, bad */
2033 irg = current_ir_graph;
2034 for (i = 0; i < n; ++i) {
2035 pred = get_irn_n(phi, i);
2036 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2038 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2039 } /* apply_binop_on_phi */
2042 * Apply an evaluator on a binop with two constant Phi.
2044 * @param a the left Phi node
2045 * @param b the right Phi node
2046 * @param eval an evaluator function
2047 * @param mode the mode of the result, may be different from the mode of the Phi!
2049 * @return a new Phi node if the conversion was successful, NULL else
2051 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, eval_func eval, ir_mode *mode) {
2052 tarval *tv_l, *tv_r, *tv;
2058 if (get_nodes_block(a) != get_nodes_block(b))
2061 n = get_irn_arity(a);
2062 NEW_ARR_A(void *, res, n);
2064 for (i = 0; i < n; ++i) {
2065 pred = get_irn_n(a, i);
2066 tv_l = get_Const_tarval(pred);
2067 pred = get_irn_n(b, i);
2068 tv_r = get_Const_tarval(pred);
2069 tv = do_eval(eval, tv_l, tv_r, mode);
2071 if (tv == tarval_bad) {
2072 /* folding failed, bad */
2077 irg = current_ir_graph;
2078 for (i = 0; i < n; ++i) {
2079 pred = get_irn_n(a, i);
2080 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2082 return new_r_Phi(get_nodes_block(a), n, (ir_node **)res, mode);
2083 } /* apply_binop_on_2_phis */
2086 * Apply an evaluator on a unop with a constant operator (a Phi).
2088 * @param phi the Phi node
2089 * @param eval an evaluator function
2091 * @return a new Phi node if the conversion was successful, NULL else
2093 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
2099 int i, n = get_irn_arity(phi);
2101 NEW_ARR_A(void *, res, n);
2102 for (i = 0; i < n; ++i) {
2103 pred = get_irn_n(phi, i);
2104 tv = get_Const_tarval(pred);
2107 if (tv == tarval_bad) {
2108 /* folding failed, bad */
2113 mode = get_irn_mode(phi);
2114 irg = current_ir_graph;
2115 for (i = 0; i < n; ++i) {
2116 pred = get_irn_n(phi, i);
2117 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2119 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2120 } /* apply_unop_on_phi */
2123 * Apply a conversion on a constant operator (a Phi).
2125 * @param phi the Phi node
2127 * @return a new Phi node if the conversion was successful, NULL else
2129 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
2134 int i, n = get_irn_arity(phi);
2136 NEW_ARR_A(void *, res, n);
2137 for (i = 0; i < n; ++i) {
2138 pred = get_irn_n(phi, i);
2139 tv = get_Const_tarval(pred);
2140 tv = tarval_convert_to(tv, mode);
2142 if (tv == tarval_bad) {
2143 /* folding failed, bad */
2148 irg = current_ir_graph;
2149 for (i = 0; i < n; ++i) {
2150 pred = get_irn_n(phi, i);
2151 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2153 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2154 } /* apply_conv_on_phi */
2157 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
2158 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
2159 * If possible, remove the Conv's.
2161 static ir_node *transform_node_AddSub(ir_node *n) {
2162 ir_mode *mode = get_irn_mode(n);
2164 if (mode_is_reference(mode)) {
2165 ir_node *left = get_binop_left(n);
2166 ir_node *right = get_binop_right(n);
2167 unsigned ref_bits = get_mode_size_bits(mode);
2169 if (is_Conv(left)) {
2170 ir_mode *lmode = get_irn_mode(left);
2171 unsigned bits = get_mode_size_bits(lmode);
2173 if (ref_bits == bits &&
2174 mode_is_int(lmode) &&
2175 get_mode_arithmetic(lmode) == irma_twos_complement) {
2176 ir_node *pre = get_Conv_op(left);
2177 ir_mode *pre_mode = get_irn_mode(pre);
2179 if (mode_is_int(pre_mode) &&
2180 get_mode_size_bits(pre_mode) == bits &&
2181 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2182 /* ok, this conv just changes to sign, moreover the calculation
2183 * is done with same number of bits as our address mode, so
2184 * we can ignore the conv as address calculation can be viewed
2185 * as either signed or unsigned
2187 set_binop_left(n, pre);
2192 if (is_Conv(right)) {
2193 ir_mode *rmode = get_irn_mode(right);
2194 unsigned bits = get_mode_size_bits(rmode);
2196 if (ref_bits == bits &&
2197 mode_is_int(rmode) &&
2198 get_mode_arithmetic(rmode) == irma_twos_complement) {
2199 ir_node *pre = get_Conv_op(right);
2200 ir_mode *pre_mode = get_irn_mode(pre);
2202 if (mode_is_int(pre_mode) &&
2203 get_mode_size_bits(pre_mode) == bits &&
2204 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2205 /* ok, this conv just changes to sign, moreover the calculation
2206 * is done with same number of bits as our address mode, so
2207 * we can ignore the conv as address calculation can be viewed
2208 * as either signed or unsigned
2210 set_binop_right(n, pre);
2215 /* let address arithmetic use unsigned modes */
2216 if (is_Const(right)) {
2217 ir_mode *rmode = get_irn_mode(right);
2219 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2220 /* convert a AddP(P, *s) into AddP(P, *u) */
2221 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2223 ir_node *pre = new_r_Conv(get_nodes_block(n), right, nm);
2224 set_binop_right(n, pre);
2230 } /* transform_node_AddSub */
2232 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2234 if (is_Const(b) && is_const_Phi(a)) { \
2235 /* check for Op(Phi, Const) */ \
2236 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2238 else if (is_Const(a) && is_const_Phi(b)) { \
2239 /* check for Op(Const, Phi) */ \
2240 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2242 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2243 /* check for Op(Phi, Phi) */ \
2244 c = apply_binop_on_2_phis(a, b, eval, mode); \
2247 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2251 #define HANDLE_UNOP_PHI(eval, a, c) \
2253 if (is_const_Phi(a)) { \
2254 /* check for Op(Phi) */ \
2255 c = apply_unop_on_phi(a, eval); \
2257 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2263 * Do the AddSub optimization, then Transform
2264 * Constant folding on Phi
2265 * Add(a,a) -> Mul(a, 2)
2266 * Add(Mul(a, x), a) -> Mul(a, x+1)
2267 * if the mode is integer or float.
2268 * Transform Add(a,-b) into Sub(a,b).
2269 * Reassociation might fold this further.
2271 static ir_node *transform_node_Add(ir_node *n) {
2273 ir_node *a, *b, *c, *oldn = n;
2275 n = transform_node_AddSub(n);
2277 a = get_Add_left(n);
2278 b = get_Add_right(n);
2280 mode = get_irn_mode(n);
2282 if (mode_is_reference(mode)) {
2283 ir_mode *lmode = get_irn_mode(a);
2285 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2286 /* an Add(a, NULL) is a hidden Conv */
2287 dbg_info *dbg = get_irn_dbg_info(n);
2288 return new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2292 HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode);
2294 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2295 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2298 if (mode_is_num(mode)) {
2299 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2300 if (!is_irg_state(current_ir_graph, IR_GRAPH_STATE_ARCH_DEP)
2301 && a == b && mode_is_int(mode)) {
2302 ir_node *block = get_nodes_block(n);
2305 get_irn_dbg_info(n),
2308 new_Const_long(mode, 2),
2310 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2315 get_irn_dbg_info(n),
2320 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2325 get_irn_dbg_info(n),
2330 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2333 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2334 /* Here we rely on constants be on the RIGHT side */
2336 ir_node *op = get_Not_op(a);
2338 if (is_Const(b) && is_Const_one(b)) {
2340 ir_node *blk = get_nodes_block(n);
2341 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, mode);
2342 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2347 n = new_Const(get_mode_minus_one(mode));
2348 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2353 ir_node *op = get_Not_op(b);
2357 n = new_Const(get_mode_minus_one(mode));
2358 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2364 if (mode_is_int(mode)) {
2365 tarval *c = tarval_and(
2366 tarval_not(a->vrp.bits_not_set),
2367 tarval_not(b->vrp.bits_not_set)
2370 if(tarval_is_null(c)) {
2371 dbg_info *dbgi = get_irn_dbg_info(n);
2372 return new_rd_Or(dbgi, get_nodes_block(n),
2377 } /* transform_node_Add */
2380 * returns -cnst or NULL if impossible
2382 static ir_node *const_negate(ir_node *cnst) {
2383 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2384 dbg_info *dbgi = get_irn_dbg_info(cnst);
2385 ir_graph *irg = get_irn_irg(cnst);
2386 if (tv == tarval_bad) return NULL;
2387 return new_rd_Const(dbgi, irg, tv);
2391 * Do the AddSub optimization, then Transform
2392 * Constant folding on Phi
2393 * Sub(0,a) -> Minus(a)
2394 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2395 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2396 * Sub(Add(a, x), x) -> a
2397 * Sub(x, Add(x, a)) -> -a
2398 * Sub(x, Const) -> Add(x, -Const)
2400 static ir_node *transform_node_Sub(ir_node *n) {
2405 n = transform_node_AddSub(n);
2407 a = get_Sub_left(n);
2408 b = get_Sub_right(n);
2410 mode = get_irn_mode(n);
2412 if (mode_is_int(mode)) {
2413 ir_mode *lmode = get_irn_mode(a);
2415 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2416 /* a Sub(a, NULL) is a hidden Conv */
2417 dbg_info *dbg = get_irn_dbg_info(n);
2418 n = new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2419 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2423 if (mode == lmode &&
2424 get_mode_arithmetic(mode) == irma_twos_complement &&
2426 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2428 dbg_info *dbg = get_irn_dbg_info(n);
2429 n = new_rd_Not(dbg, get_nodes_block(n), b, mode);
2430 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2436 HANDLE_BINOP_PHI((eval_func) tarval_sub, a, b, c, mode);
2438 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2439 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2442 if (is_Const(b) && !mode_is_reference(get_irn_mode(b))) {
2443 /* a - C -> a + (-C) */
2444 ir_node *cnst = const_negate(b);
2446 ir_node *block = get_nodes_block(n);
2447 dbg_info *dbgi = get_irn_dbg_info(n);
2449 n = new_rd_Add(dbgi, block, a, cnst, mode);
2450 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2455 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2456 dbg_info *dbg = get_irn_dbg_info(n);
2457 ir_node *block = get_nodes_block(n);
2458 ir_node *left = get_Minus_op(a);
2459 ir_node *add = new_rd_Add(dbg, block, left, b, mode);
2461 n = new_rd_Minus(dbg, block, add, mode);
2462 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2464 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2465 dbg_info *dbg = get_irn_dbg_info(n);
2466 ir_node *block = get_nodes_block(n);
2467 ir_node *right = get_Minus_op(b);
2469 n = new_rd_Add(dbg, block, a, right, mode);
2470 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2472 } else if (is_Sub(b)) {
2473 /* a - (b - c) -> a + (c - b)
2474 * -> (a - b) + c iff (b - c) is a pointer */
2475 dbg_info *s_dbg = get_irn_dbg_info(b);
2476 ir_node *s_block = get_nodes_block(b);
2477 ir_node *s_left = get_Sub_left(b);
2478 ir_node *s_right = get_Sub_right(b);
2479 ir_mode *s_mode = get_irn_mode(b);
2480 if (mode_is_reference(s_mode)) {
2481 ir_node *sub = new_rd_Sub(s_dbg, s_block, a, s_left, mode);
2482 dbg_info *a_dbg = get_irn_dbg_info(n);
2483 ir_node *a_block = get_nodes_block(n);
2486 s_right = new_r_Conv(a_block, s_right, mode);
2487 n = new_rd_Add(a_dbg, a_block, sub, s_right, mode);
2489 ir_node *sub = new_rd_Sub(s_dbg, s_block, s_right, s_left, s_mode);
2490 dbg_info *a_dbg = get_irn_dbg_info(n);
2491 ir_node *a_block = get_nodes_block(n);
2493 n = new_rd_Add(a_dbg, a_block, a, sub, mode);
2495 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2497 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2498 ir_node *m_right = get_Mul_right(b);
2499 if (is_Const(m_right)) {
2500 ir_node *cnst2 = const_negate(m_right);
2501 if (cnst2 != NULL) {
2502 dbg_info *m_dbg = get_irn_dbg_info(b);
2503 ir_node *m_block = get_nodes_block(b);
2504 ir_node *m_left = get_Mul_left(b);
2505 ir_mode *m_mode = get_irn_mode(b);
2506 ir_node *mul = new_rd_Mul(m_dbg, m_block, m_left, cnst2, m_mode);
2507 dbg_info *a_dbg = get_irn_dbg_info(n);
2508 ir_node *a_block = get_nodes_block(n);
2510 n = new_rd_Add(a_dbg, a_block, a, mul, mode);
2511 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2517 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2518 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2520 get_irn_dbg_info(n),
2524 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2528 if (mode_wrap_around(mode)) {
2529 ir_node *left = get_Add_left(a);
2530 ir_node *right = get_Add_right(a);
2532 /* FIXME: Does the Conv's work only for two complement or generally? */
2534 if (mode != get_irn_mode(right)) {
2535 /* This Sub is an effective Cast */
2536 right = new_r_Conv(get_nodes_block(n), right, mode);
2539 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2541 } else if (right == b) {
2542 if (mode != get_irn_mode(left)) {
2543 /* This Sub is an effective Cast */
2544 left = new_r_Conv(get_nodes_block(n), left, mode);
2547 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2553 if (mode_wrap_around(mode)) {
2554 ir_node *left = get_Add_left(b);
2555 ir_node *right = get_Add_right(b);
2557 /* FIXME: Does the Conv's work only for two complement or generally? */
2559 ir_mode *r_mode = get_irn_mode(right);
2561 n = new_r_Minus(get_nodes_block(n), right, r_mode);
2562 if (mode != r_mode) {
2563 /* This Sub is an effective Cast */
2564 n = new_r_Conv(get_nodes_block(n), n, mode);
2566 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2568 } else if (right == a) {
2569 ir_mode *l_mode = get_irn_mode(left);
2571 n = new_r_Minus(get_nodes_block(n), left, l_mode);
2572 if (mode != l_mode) {
2573 /* This Sub is an effective Cast */
2574 n = new_r_Conv(get_nodes_block(n), n, mode);
2576 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2581 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2582 ir_mode *mode = get_irn_mode(a);
2584 if (mode == get_irn_mode(b)) {
2586 ir_node *op_a = get_Conv_op(a);
2587 ir_node *op_b = get_Conv_op(b);
2589 /* check if it's allowed to skip the conv */
2590 ma = get_irn_mode(op_a);
2591 mb = get_irn_mode(op_b);
2593 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2594 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2597 set_Sub_right(n, b);
2603 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2604 if (!is_reassoc_running() && is_Mul(a)) {
2605 ir_node *ma = get_Mul_left(a);
2606 ir_node *mb = get_Mul_right(a);
2609 ir_node *blk = get_nodes_block(n);
2611 get_irn_dbg_info(n),
2615 get_irn_dbg_info(n),
2618 new_Const_long(mode, 1),
2621 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2623 } else if (mb == b) {
2624 ir_node *blk = get_nodes_block(n);
2626 get_irn_dbg_info(n),
2630 get_irn_dbg_info(n),
2633 new_Const_long(mode, 1),
2636 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2640 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2641 ir_node *x = get_Sub_left(a);
2642 ir_node *y = get_Sub_right(a);
2643 ir_node *blk = get_nodes_block(n);
2644 ir_mode *m_b = get_irn_mode(b);
2645 ir_mode *m_y = get_irn_mode(y);
2649 /* Determine the right mode for the Add. */
2652 else if (mode_is_reference(m_b))
2654 else if (mode_is_reference(m_y))
2658 * Both modes are different but none is reference,
2659 * happens for instance in SubP(SubP(P, Iu), Is).
2660 * We have two possibilities here: Cast or ignore.
2661 * Currently we ignore this case.
2666 add = new_r_Add(blk, y, b, add_mode);
2668 n = new_rd_Sub(get_irn_dbg_info(n), blk, x, add, mode);
2669 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2673 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2674 if (is_Const(a) && is_Not(b)) {
2675 /* c - ~X = X + (c+1) */
2676 tarval *tv = get_Const_tarval(a);
2678 tv = tarval_add(tv, get_mode_one(mode));
2679 if (tv != tarval_bad) {
2680 ir_node *blk = get_nodes_block(n);
2681 ir_node *c = new_Const(tv);
2682 n = new_rd_Add(get_irn_dbg_info(n), blk, get_Not_op(b), c, mode);
2683 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2689 } /* transform_node_Sub */
2692 * Several transformation done on n*n=2n bits mul.
2693 * These transformations must be done here because new nodes may be produced.
2695 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2697 ir_node *a = get_Mul_left(n);
2698 ir_node *b = get_Mul_right(n);
2699 tarval *ta = value_of(a);
2700 tarval *tb = value_of(b);
2701 ir_mode *smode = get_irn_mode(a);
2703 if (ta == get_mode_one(smode)) {
2704 /* (L)1 * (L)b = (L)b */
2705 ir_node *blk = get_nodes_block(n);
2706 n = new_rd_Conv(get_irn_dbg_info(n), blk, b, mode);
2707 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2710 else if (ta == get_mode_minus_one(smode)) {
2711 /* (L)-1 * (L)b = (L)b */
2712 ir_node *blk = get_nodes_block(n);
2713 n = new_rd_Minus(get_irn_dbg_info(n), blk, b, smode);
2714 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2715 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2718 if (tb == get_mode_one(smode)) {
2719 /* (L)a * (L)1 = (L)a */
2720 ir_node *blk = get_irn_n(a, -1);
2721 n = new_rd_Conv(get_irn_dbg_info(n), blk, a, mode);
2722 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2725 else if (tb == get_mode_minus_one(smode)) {
2726 /* (L)a * (L)-1 = (L)-a */
2727 ir_node *blk = get_nodes_block(n);
2728 n = new_rd_Minus(get_irn_dbg_info(n), blk, a, smode);
2729 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2730 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2737 * Transform Mul(a,-1) into -a.
2738 * Do constant evaluation of Phi nodes.
2739 * Do architecture dependent optimizations on Mul nodes
2741 static ir_node *transform_node_Mul(ir_node *n) {
2742 ir_node *c, *oldn = n;
2743 ir_mode *mode = get_irn_mode(n);
2744 ir_node *a = get_Mul_left(n);
2745 ir_node *b = get_Mul_right(n);
2747 if (is_Bad(a) || is_Bad(b))
2750 if (mode != get_irn_mode(a))
2751 return transform_node_Mul2n(n, mode);
2753 HANDLE_BINOP_PHI((eval_func) tarval_mul, a, b, c, mode);
2755 if (mode_is_signed(mode)) {
2758 if (value_of(a) == get_mode_minus_one(mode))
2760 else if (value_of(b) == get_mode_minus_one(mode))
2763 n = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), r, mode);
2764 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2769 if (is_Const(b)) { /* (-a) * const -> a * -const */
2770 ir_node *cnst = const_negate(b);
2772 dbg_info *dbgi = get_irn_dbg_info(n);
2773 ir_node *block = get_nodes_block(n);
2774 n = new_rd_Mul(dbgi, block, get_Minus_op(a), cnst, mode);
2775 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2778 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2779 dbg_info *dbgi = get_irn_dbg_info(n);
2780 ir_node *block = get_nodes_block(n);
2781 n = new_rd_Mul(dbgi, block, get_Minus_op(a), get_Minus_op(b), mode);
2782 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2784 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2785 ir_node *sub_l = get_Sub_left(b);
2786 ir_node *sub_r = get_Sub_right(b);
2787 dbg_info *dbgi = get_irn_dbg_info(n);
2788 ir_node *block = get_nodes_block(n);
2789 ir_node *new_b = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2790 n = new_rd_Mul(dbgi, block, get_Minus_op(a), new_b, mode);
2791 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2794 } else if (is_Minus(b)) {
2795 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2796 ir_node *sub_l = get_Sub_left(a);
2797 ir_node *sub_r = get_Sub_right(a);
2798 dbg_info *dbgi = get_irn_dbg_info(n);
2799 ir_node *block = get_nodes_block(n);
2800 ir_node *new_a = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2801 n = new_rd_Mul(dbgi, block, new_a, get_Minus_op(b), mode);
2802 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2805 } else if (is_Shl(a)) {
2806 ir_node *const shl_l = get_Shl_left(a);
2807 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2808 /* (1 << x) * b -> b << x */
2809 dbg_info *const dbgi = get_irn_dbg_info(n);
2810 ir_node *const block = get_nodes_block(n);
2811 ir_node *const shl_r = get_Shl_right(a);
2812 n = new_rd_Shl(dbgi, block, b, shl_r, mode);
2813 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2816 } else if (is_Shl(b)) {
2817 ir_node *const shl_l = get_Shl_left(b);
2818 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2819 /* a * (1 << x) -> a << x */
2820 dbg_info *const dbgi = get_irn_dbg_info(n);
2821 ir_node *const block = get_nodes_block(n);
2822 ir_node *const shl_r = get_Shl_right(b);
2823 n = new_rd_Shl(dbgi, block, a, shl_r, mode);
2824 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2828 if (get_mode_arithmetic(mode) == irma_ieee754) {
2830 tarval *tv = get_Const_tarval(a);
2831 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2832 && !tarval_is_negative(tv)) {
2833 /* 2.0 * b = b + b */
2834 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), b, b, mode);
2835 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2839 else if (is_Const(b)) {
2840 tarval *tv = get_Const_tarval(b);
2841 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2842 && !tarval_is_negative(tv)) {
2843 /* a * 2.0 = a + a */
2844 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), a, a, mode);
2845 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2850 return arch_dep_replace_mul_with_shifts(n);
2851 } /* transform_node_Mul */
2854 * Transform a Div Node.
2856 static ir_node *transform_node_Div(ir_node *n) {
2857 ir_mode *mode = get_Div_resmode(n);
2858 ir_node *a = get_Div_left(n);
2859 ir_node *b = get_Div_right(n);
2861 const ir_node *dummy;
2863 if (is_Const(b) && is_const_Phi(a)) {
2864 /* check for Div(Phi, Const) */
2865 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0);
2867 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2871 else if (is_Const(a) && is_const_Phi(b)) {
2872 /* check for Div(Const, Phi) */
2873 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1);
2875 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2879 else if (is_const_Phi(a) && is_const_Phi(b)) {
2880 /* check for Div(Phi, Phi) */
2881 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode);
2883 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2890 if (a == b && value_not_zero(a, &dummy)) {
2891 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2892 value = new_Const(get_mode_one(mode));
2893 DBG_OPT_CSTEVAL(n, value);
2896 if (mode_is_signed(mode) && is_Const(b)) {
2897 tarval *tv = get_Const_tarval(b);
2899 if (tv == get_mode_minus_one(mode)) {
2901 value = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
2902 DBG_OPT_CSTEVAL(n, value);
2906 /* Try architecture dependent optimization */
2907 value = arch_dep_replace_div_by_const(n);
2914 /* Turn Div into a tuple (mem, jmp, bad, value) */
2915 mem = get_Div_mem(n);
2916 blk = get_nodes_block(n);
2918 /* skip a potential Pin */
2919 mem = skip_Pin(mem);
2920 turn_into_tuple(n, pn_Div_max);
2921 set_Tuple_pred(n, pn_Div_M, mem);
2922 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(blk));
2923 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2924 set_Tuple_pred(n, pn_Div_res, value);
2927 } /* transform_node_Div */
2930 * Transform a Mod node.
2932 static ir_node *transform_node_Mod(ir_node *n) {
2933 ir_mode *mode = get_Mod_resmode(n);
2934 ir_node *a = get_Mod_left(n);
2935 ir_node *b = get_Mod_right(n);
2939 if (is_Const(b) && is_const_Phi(a)) {
2940 /* check for Div(Phi, Const) */
2941 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0);
2943 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2947 else if (is_Const(a) && is_const_Phi(b)) {
2948 /* check for Div(Const, Phi) */
2949 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1);
2951 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2955 else if (is_const_Phi(a) && is_const_Phi(b)) {
2956 /* check for Div(Phi, Phi) */
2957 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode);
2959 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2966 if (tv != tarval_bad) {
2967 value = new_Const(tv);
2969 DBG_OPT_CSTEVAL(n, value);
2972 ir_node *a = get_Mod_left(n);
2973 ir_node *b = get_Mod_right(n);
2974 const ir_node *dummy;
2976 if (a == b && value_not_zero(a, &dummy)) {
2977 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2978 value = new_Const(get_mode_null(mode));
2979 DBG_OPT_CSTEVAL(n, value);
2982 if (mode_is_signed(mode) && is_Const(b)) {
2983 tarval *tv = get_Const_tarval(b);
2985 if (tv == get_mode_minus_one(mode)) {
2987 value = new_Const(get_mode_null(mode));
2988 DBG_OPT_CSTEVAL(n, value);
2992 /* Try architecture dependent optimization */
2993 value = arch_dep_replace_mod_by_const(n);
3001 /* Turn Mod into a tuple (mem, jmp, bad, value) */
3002 mem = get_Mod_mem(n);
3003 blk = get_nodes_block(n);
3005 /* skip a potential Pin */
3006 mem = skip_Pin(mem);
3007 turn_into_tuple(n, pn_Mod_max);
3008 set_Tuple_pred(n, pn_Mod_M, mem);
3009 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(blk));
3010 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
3011 set_Tuple_pred(n, pn_Mod_res, value);
3014 } /* transform_node_Mod */
3017 * Transform a DivMod node.
3019 static ir_node *transform_node_DivMod(ir_node *n) {
3020 const ir_node *dummy;
3021 ir_node *a = get_DivMod_left(n);
3022 ir_node *b = get_DivMod_right(n);
3023 ir_mode *mode = get_DivMod_resmode(n);
3028 if (is_Const(b) && is_const_Phi(a)) {
3029 /* check for Div(Phi, Const) */
3030 va = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0);
3031 vb = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0);
3033 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
3034 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
3038 else if (is_Const(a) && is_const_Phi(b)) {
3039 /* check for Div(Const, Phi) */
3040 va = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1);
3041 vb = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1);
3043 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
3044 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
3048 else if (is_const_Phi(a) && is_const_Phi(b)) {
3049 /* check for Div(Phi, Phi) */
3050 va = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode);
3051 vb = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode);
3053 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
3054 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
3061 if (tb != tarval_bad) {
3062 if (tb == get_mode_one(get_tarval_mode(tb))) {
3064 vb = new_Const(get_mode_null(mode));
3065 DBG_OPT_CSTEVAL(n, vb);
3067 } else if (ta != tarval_bad) {
3068 tarval *resa, *resb;
3069 resa = tarval_div(ta, tb);
3070 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
3071 Jmp for X result!? */
3072 resb = tarval_mod(ta, tb);
3073 if (resb == tarval_bad) return n; /* Causes exception! */
3074 va = new_Const(resa);
3075 vb = new_Const(resb);
3076 DBG_OPT_CSTEVAL(n, va);
3077 DBG_OPT_CSTEVAL(n, vb);
3079 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
3080 va = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
3081 vb = new_Const(get_mode_null(mode));
3082 DBG_OPT_CSTEVAL(n, va);
3083 DBG_OPT_CSTEVAL(n, vb);
3085 } else { /* Try architecture dependent optimization */
3088 arch_dep_replace_divmod_by_const(&va, &vb, n);
3089 evaluated = va != NULL;
3091 } else if (a == b) {
3092 if (value_not_zero(a, &dummy)) {
3094 va = new_Const(get_mode_one(mode));
3095 vb = new_Const(get_mode_null(mode));
3096 DBG_OPT_CSTEVAL(n, va);
3097 DBG_OPT_CSTEVAL(n, vb);
3100 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
3103 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
3104 /* 0 / non-Const = 0 */
3109 if (evaluated) { /* replace by tuple */
3113 mem = get_DivMod_mem(n);
3114 /* skip a potential Pin */
3115 mem = skip_Pin(mem);
3117 blk = get_nodes_block(n);
3118 turn_into_tuple(n, pn_DivMod_max);
3119 set_Tuple_pred(n, pn_DivMod_M, mem);
3120 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(blk));
3121 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
3122 set_Tuple_pred(n, pn_DivMod_res_div, va);
3123 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
3127 } /* transform_node_DivMod */
3130 * Optimize x / c to x * (1/c)
3132 static ir_node *transform_node_Quot(ir_node *n) {
3133 ir_mode *mode = get_Quot_resmode(n);
3136 if (get_mode_arithmetic(mode) == irma_ieee754) {
3137 ir_node *b = get_Quot_right(n);
3138 tarval *tv = value_of(b);
3140 if (tv != tarval_bad) {
3141 int rem = tarval_fp_ops_enabled();
3144 * Floating point constant folding might be disabled here to
3146 * However, as we check for exact result, doing it is safe.
3149 tarval_enable_fp_ops(1);
3150 tv = tarval_quo(get_mode_one(mode), tv);
3151 tarval_enable_fp_ops(rem);
3153 /* Do the transformation if the result is either exact or we are not
3154 using strict rules. */
3155 if (tv != tarval_bad &&
3156 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
3157 ir_node *blk = get_nodes_block(n);
3158 ir_node *c = new_Const(tv);
3159 ir_node *a = get_Quot_left(n);
3160 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), blk, a, c, mode);
3161 ir_node *mem = get_Quot_mem(n);
3163 /* skip a potential Pin */
3164 mem = skip_Pin(mem);
3165 turn_into_tuple(n, pn_Quot_max);
3166 set_Tuple_pred(n, pn_Quot_M, mem);
3167 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(blk));
3168 set_Tuple_pred(n, pn_Quot_X_except, new_Bad());
3169 set_Tuple_pred(n, pn_Quot_res, m);
3170 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
3175 } /* transform_node_Quot */
3178 * Optimize Abs(x) into x if x is Confirmed >= 0
3179 * Optimize Abs(x) into -x if x is Confirmed <= 0
3180 * Optimize Abs(-x) int Abs(x)
3182 static ir_node *transform_node_Abs(ir_node *n) {
3183 ir_node *c, *oldn = n;
3184 ir_node *a = get_Abs_op(n);
3187 HANDLE_UNOP_PHI(tarval_abs, a, c);
3189 switch (classify_value_sign(a)) {
3190 case value_classified_negative:
3191 mode = get_irn_mode(n);
3194 * We can replace the Abs by -x here.
3195 * We even could add a new Confirm here
3196 * (if not twos complement)
3198 * Note that -x would create a new node, so we could
3199 * not run it in the equivalent_node() context.
3201 n = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
3203 DBG_OPT_CONFIRM(oldn, n);
3205 case value_classified_positive:
3206 /* n is positive, Abs is not needed */
3209 DBG_OPT_CONFIRM(oldn, n);
3215 /* Abs(-x) = Abs(x) */
3216 mode = get_irn_mode(n);
3217 n = new_rd_Abs(get_irn_dbg_info(n), get_nodes_block(n), get_Minus_op(a), mode);
3218 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
3222 } /* transform_node_Abs */
3225 * Optimize -a CMP -b into b CMP a.
3226 * This works only for for modes where unary Minus
3228 * Note that two-complement integers can Overflow
3229 * so it will NOT work.
3231 * For == and != can be handled in Proj(Cmp)
3233 static ir_node *transform_node_Cmp(ir_node *n) {
3235 ir_node *left = get_Cmp_left(n);
3236 ir_node *right = get_Cmp_right(n);
3238 if (is_Minus(left) && is_Minus(right) &&
3239 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
3240 ir_node *const new_left = get_Minus_op(right);
3241 ir_node *const new_right = get_Minus_op(left);
3242 n = new_rd_Cmp(get_irn_dbg_info(n), get_nodes_block(n), new_left, new_right);
3243 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP);
3246 } /* transform_node_Cmp */
3250 * Transform a Cond node.
3252 * Replace the Cond by a Jmp if it branches on a constant
3255 static ir_node *transform_node_Cond(ir_node *n) {
3258 ir_node *a = get_Cond_selector(n);
3259 tarval *ta = value_of(a);
3261 /* we need block info which is not available in floating irgs */
3262 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
3265 if ((ta != tarval_bad) &&
3266 (get_irn_mode(a) == mode_b) &&
3267 (get_opt_unreachable_code())) {
3268 /* It's a boolean Cond, branching on a boolean constant.
3269 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3270 ir_node *blk = get_nodes_block(n);
3271 jmp = new_r_Jmp(blk);
3272 turn_into_tuple(n, pn_Cond_max);
3273 if (ta == tarval_b_true) {
3274 set_Tuple_pred(n, pn_Cond_false, new_Bad());
3275 set_Tuple_pred(n, pn_Cond_true, jmp);
3277 set_Tuple_pred(n, pn_Cond_false, jmp);
3278 set_Tuple_pred(n, pn_Cond_true, new_Bad());
3280 /* We might generate an endless loop, so keep it alive. */
3281 add_End_keepalive(get_irg_end(current_ir_graph), blk);
3284 } /* transform_node_Cond */
3287 * Prototype of a recursive transform function
3288 * for bitwise distributive transformations.
3290 typedef ir_node* (*recursive_transform)(ir_node *n);
3293 * makes use of distributive laws for and, or, eor
3294 * and(a OP c, b OP c) -> and(a, b) OP c
3295 * note, might return a different op than n
3297 static ir_node *transform_bitwise_distributive(ir_node *n,
3298 recursive_transform trans_func)
3301 ir_node *a = get_binop_left(n);
3302 ir_node *b = get_binop_right(n);
3303 ir_op *op = get_irn_op(a);
3304 ir_op *op_root = get_irn_op(n);
3306 if(op != get_irn_op(b))
3309 /* and(conv(a), conv(b)) -> conv(and(a,b)) */
3310 if (op == op_Conv) {
3311 ir_node *a_op = get_Conv_op(a);
3312 ir_node *b_op = get_Conv_op(b);
3313 ir_mode *a_mode = get_irn_mode(a_op);
3314 ir_mode *b_mode = get_irn_mode(b_op);
3315 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3316 ir_node *blk = get_nodes_block(n);
3319 set_binop_left(n, a_op);
3320 set_binop_right(n, b_op);
3321 set_irn_mode(n, a_mode);
3323 n = new_r_Conv(blk, n, get_irn_mode(oldn));
3325 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
3331 /* nothing to gain here */
3335 if (op == op_Shrs || op == op_Shr || op == op_Shl
3336 || op == op_And || op == op_Or || op == op_Eor) {
3337 ir_node *a_left = get_binop_left(a);
3338 ir_node *a_right = get_binop_right(a);
3339 ir_node *b_left = get_binop_left(b);
3340 ir_node *b_right = get_binop_right(b);
3342 ir_node *op1 = NULL;
3343 ir_node *op2 = NULL;
3345 if (is_op_commutative(op)) {
3346 if (a_left == b_left) {
3350 } else if(a_left == b_right) {
3354 } else if(a_right == b_left) {
3360 if(a_right == b_right) {
3367 /* (a sop c) & (b sop c) => (a & b) sop c */
3368 ir_node *blk = get_nodes_block(n);
3370 ir_node *new_n = exact_copy(n);
3371 set_binop_left(new_n, op1);
3372 set_binop_right(new_n, op2);
3373 new_n = trans_func(new_n);
3375 if(op_root == op_Eor && op == op_Or) {
3376 dbg_info *dbgi = get_irn_dbg_info(n);
3377 ir_mode *mode = get_irn_mode(c);
3379 c = new_rd_Not(dbgi, blk, c, mode);
3380 n = new_rd_And(dbgi, blk, new_n, c, mode);
3383 set_nodes_block(n, blk);
3384 set_binop_left(n, new_n);
3385 set_binop_right(n, c);
3386 add_identities(current_ir_graph->value_table, n);
3389 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3400 static ir_node *transform_node_And(ir_node *n) {
3401 ir_node *c, *oldn = n;
3402 ir_node *a = get_And_left(n);
3403 ir_node *b = get_And_right(n);
3406 mode = get_irn_mode(n);
3407 HANDLE_BINOP_PHI((eval_func) tarval_and, a, b, c, mode);
3409 /* we can evaluate 2 Projs of the same Cmp */
3410 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3411 ir_node *pred_a = get_Proj_pred(a);
3412 ir_node *pred_b = get_Proj_pred(b);
3413 if (pred_a == pred_b) {
3414 dbg_info *dbgi = get_irn_dbg_info(n);
3415 ir_node *block = get_nodes_block(pred_a);
3416 pn_Cmp pn_a = get_Proj_proj(a);
3417 pn_Cmp pn_b = get_Proj_proj(b);
3418 /* yes, we can simply calculate with pncs */
3419 pn_Cmp new_pnc = pn_a & pn_b;
3421 return new_rd_Proj(dbgi, block, pred_a, mode_b, new_pnc);
3426 ir_node *op = get_Not_op(b);
3428 ir_node *ba = get_And_left(op);
3429 ir_node *bb = get_And_right(op);
3431 /* it's enough to test the following cases due to normalization! */
3432 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3433 /* (a|b) & ~(a&b) = a^b */
3434 ir_node *block = get_nodes_block(n);
3436 n = new_rd_Eor(get_irn_dbg_info(n), block, ba, bb, mode);
3437 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3445 ir_node *op = get_Not_op(a);
3447 ir_node *aa = get_And_left(op);
3448 ir_node *ab = get_And_right(op);
3450 /* it's enough to test the following cases due to normalization! */
3451 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3452 /* (a|b) & ~(a&b) = a^b */
3453 ir_node *block = get_nodes_block(n);
3455 n = new_rd_Eor(get_irn_dbg_info(n), block, aa, ab, mode);
3456 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3463 ir_node *al = get_Eor_left(a);
3464 ir_node *ar = get_Eor_right(a);
3467 /* (b ^ a) & b -> ~a & b */
3468 dbg_info *dbg = get_irn_dbg_info(n);
3469 ir_node *block = get_nodes_block(n);
3471 ar = new_rd_Not(dbg, block, ar, mode);
3472 n = new_rd_And(dbg, block, ar, b, mode);
3473 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3477 /* (a ^ b) & b -> ~a & b */
3478 dbg_info *dbg = get_irn_dbg_info(n);
3479 ir_node *block = get_nodes_block(n);
3481 al = new_rd_Not(dbg, block, al, mode);
3482 n = new_rd_And(dbg, block, al, b, mode);
3483 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3488 ir_node *bl = get_Eor_left(b);
3489 ir_node *br = get_Eor_right(b);
3492 /* a & (a ^ b) -> a & ~b */
3493 dbg_info *dbg = get_irn_dbg_info(n);
3494 ir_node *block = get_nodes_block(n);
3496 br = new_rd_Not(dbg, block, br, mode);
3497 n = new_rd_And(dbg, block, br, a, mode);
3498 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3502 /* a & (b ^ a) -> a & ~b */
3503 dbg_info *dbg = get_irn_dbg_info(n);
3504 ir_node *block = get_nodes_block(n);
3506 bl = new_rd_Not(dbg, block, bl, mode);
3507 n = new_rd_And(dbg, block, bl, a, mode);
3508 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3512 if (is_Not(a) && is_Not(b)) {
3513 /* ~a & ~b = ~(a|b) */
3514 ir_node *block = get_nodes_block(n);
3515 ir_mode *mode = get_irn_mode(n);
3519 n = new_rd_Or(get_irn_dbg_info(n), block, a, b, mode);
3520 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
3521 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3525 if (is_Const(a) && (tarval_is_all_one(tarval_or(get_Const_tarval(a),
3526 b->vrp.bits_not_set)))) {
3527 return new_rd_Id(get_irn_dbg_info(n), get_nodes_block(n),
3528 b, get_irn_mode(n));
3532 if (is_Const(b) && (tarval_is_all_one(tarval_or(get_Const_tarval(b),
3533 a->vrp.bits_not_set)))) {
3534 return new_rd_Id(get_irn_dbg_info(n), get_nodes_block(n),
3535 a, get_irn_mode(n));
3539 n = transform_bitwise_distributive(n, transform_node_And);
3542 } /* transform_node_And */
3547 static ir_node *transform_node_Eor(ir_node *n) {
3548 ir_node *c, *oldn = n;
3549 ir_node *a = get_Eor_left(n);
3550 ir_node *b = get_Eor_right(n);
3551 ir_mode *mode = get_irn_mode(n);
3553 HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode);
3555 /* we can evaluate 2 Projs of the same Cmp */
3556 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3557 ir_node *pred_a = get_Proj_pred(a);
3558 ir_node *pred_b = get_Proj_pred(b);
3559 if(pred_a == pred_b) {
3560 dbg_info *dbgi = get_irn_dbg_info(n);
3561 ir_node *block = get_nodes_block(pred_a);
3562 pn_Cmp pn_a = get_Proj_proj(a);
3563 pn_Cmp pn_b = get_Proj_proj(b);
3564 /* yes, we can simply calculate with pncs */
3565 pn_Cmp new_pnc = pn_a ^ pn_b;
3567 return new_rd_Proj(dbgi, block, pred_a, mode_b, new_pnc);
3573 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph,
3574 get_mode_null(mode));
3575 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3576 } else if (is_Const(b)) {
3577 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3578 ir_node *cnst = new_Const(tarval_not(get_Const_tarval(b)));
3579 ir_node *not_op = get_Not_op(a);
3580 dbg_info *dbg = get_irn_dbg_info(n);
3581 ir_node *block = get_nodes_block(n);
3582 ir_mode *mode = get_irn_mode(n);
3583 n = new_rd_Eor(dbg, block, not_op, cnst, mode);
3585 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3586 n = new_r_Not(get_nodes_block(n), a, mode);
3587 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3590 n = transform_bitwise_distributive(n, transform_node_Eor);
3594 } /* transform_node_Eor */
3599 static ir_node *transform_node_Not(ir_node *n) {
3600 ir_node *c, *oldn = n;
3601 ir_node *a = get_Not_op(n);
3602 ir_mode *mode = get_irn_mode(n);
3604 HANDLE_UNOP_PHI(tarval_not,a,c);
3606 /* check for a boolean Not */
3607 if (mode == mode_b && is_Proj(a)) {
3608 ir_node *a_pred = get_Proj_pred(a);
3609 if (is_Cmp(a_pred)) {
3610 ir_node *cmp_block = get_nodes_block(a_pred);
3611 /* We negate a Cmp. The Cmp has the negated result anyways! */
3612 n = new_r_Proj(cmp_block, get_Proj_pred(a),
3613 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3614 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3619 ir_node *eor_b = get_Eor_right(a);
3620 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3621 ir_node *cnst = new_Const(tarval_not(get_Const_tarval(eor_b)));
3622 ir_node *eor_a = get_Eor_left(a);
3623 dbg_info *dbg = get_irn_dbg_info(n);
3624 ir_node *block = get_nodes_block(n);
3625 ir_mode *mode = get_irn_mode(n);
3626 n = new_rd_Eor(dbg, block, eor_a, cnst, mode);
3630 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3631 if (is_Minus(a)) { /* ~-x -> x + -1 */
3632 dbg_info *dbg = get_irn_dbg_info(n);
3633 ir_graph *irg = current_ir_graph;
3634 ir_node *block = get_nodes_block(n);
3635 ir_node *add_l = get_Minus_op(a);
3636 ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode));
3637 n = new_rd_Add(dbg, block, add_l, add_r, mode);
3638 } else if (is_Add(a)) {
3639 ir_node *add_r = get_Add_right(a);
3640 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3641 /* ~(x + -1) = -x */
3642 ir_node *op = get_Add_left(a);
3643 ir_node *blk = get_nodes_block(n);
3644 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, get_irn_mode(n));
3645 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3650 } /* transform_node_Not */
3653 * Transform a Minus.
3657 * -(a >>u (size-1)) = a >>s (size-1)
3658 * -(a >>s (size-1)) = a >>u (size-1)
3659 * -(a * const) -> a * -const
3661 static ir_node *transform_node_Minus(ir_node *n) {
3662 ir_node *c, *oldn = n;
3663 ir_node *a = get_Minus_op(n);
3666 HANDLE_UNOP_PHI(tarval_neg,a,c);
3668 mode = get_irn_mode(a);
3669 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3670 /* the following rules are only to twos-complement */
3673 ir_node *op = get_Not_op(a);
3674 tarval *tv = get_mode_one(mode);
3675 ir_node *blk = get_nodes_block(n);
3676 ir_node *c = new_Const(tv);
3677 n = new_rd_Add(get_irn_dbg_info(n), blk, op, c, mode);
3678 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3682 ir_node *c = get_Shr_right(a);
3685 tarval *tv = get_Const_tarval(c);
3687 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3688 /* -(a >>u (size-1)) = a >>s (size-1) */
3689 ir_node *v = get_Shr_left(a);
3691 n = new_rd_Shrs(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3692 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3698 ir_node *c = get_Shrs_right(a);
3701 tarval *tv = get_Const_tarval(c);
3703 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3704 /* -(a >>s (size-1)) = a >>u (size-1) */
3705 ir_node *v = get_Shrs_left(a);
3707 n = new_rd_Shr(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3708 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3715 /* - (a-b) = b - a */
3716 ir_node *la = get_Sub_left(a);
3717 ir_node *ra = get_Sub_right(a);
3718 ir_node *blk = get_nodes_block(n);
3720 n = new_rd_Sub(get_irn_dbg_info(n), blk, ra, la, mode);
3721 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3725 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3726 ir_node *mul_l = get_Mul_left(a);
3727 ir_node *mul_r = get_Mul_right(a);
3728 tarval *tv = value_of(mul_r);
3729 if (tv != tarval_bad) {
3730 tv = tarval_neg(tv);
3731 if (tv != tarval_bad) {
3732 ir_node *cnst = new_Const(tv);
3733 dbg_info *dbg = get_irn_dbg_info(a);
3734 ir_node *block = get_nodes_block(a);
3735 n = new_rd_Mul(dbg, block, mul_l, cnst, mode);
3736 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3743 } /* transform_node_Minus */
3746 * Transform a Cast_type(Const) into a new Const_type
3748 static ir_node *transform_node_Cast(ir_node *n) {
3750 ir_node *pred = get_Cast_op(n);
3751 ir_type *tp = get_irn_type(n);
3753 if (is_Const(pred) && get_Const_type(pred) != tp) {
3754 n = new_rd_Const_type(NULL, current_ir_graph, get_Const_tarval(pred), tp);
3755 DBG_OPT_CSTEVAL(oldn, n);
3756 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3757 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_mode(pred),
3758 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3759 DBG_OPT_CSTEVAL(oldn, n);
3763 } /* transform_node_Cast */
3766 * Transform a Proj(Load) with a non-null address.
3768 static ir_node *transform_node_Proj_Load(ir_node *proj) {
3769 if (get_opt_ldst_only_null_ptr_exceptions()) {
3770 if (get_irn_mode(proj) == mode_X) {
3771 ir_node *load = get_Proj_pred(proj);
3773 /* get the Load address */
3774 const ir_node *addr = get_Load_ptr(load);
3775 const ir_node *confirm;
3777 if (value_not_null(addr, &confirm)) {
3778 if (confirm == NULL) {
3779 /* this node may float if it did not depend on a Confirm */
3780 set_irn_pinned(load, op_pin_state_floats);
3782 if (get_Proj_proj(proj) == pn_Load_X_except) {
3783 DBG_OPT_EXC_REM(proj);
3784 return get_irg_bad(current_ir_graph);
3786 ir_node *blk = get_nodes_block(load);
3787 return new_r_Jmp(blk);
3793 } /* transform_node_Proj_Load */
3796 * Transform a Proj(Store) with a non-null address.
3798 static ir_node *transform_node_Proj_Store(ir_node *proj) {
3799 if (get_opt_ldst_only_null_ptr_exceptions()) {
3800 if (get_irn_mode(proj) == mode_X) {
3801 ir_node *store = get_Proj_pred(proj);
3803 /* get the load/store address */
3804 const ir_node *addr = get_Store_ptr(store);
3805 const ir_node *confirm;
3807 if (value_not_null(addr, &confirm)) {
3808 if (confirm == NULL) {
3809 /* this node may float if it did not depend on a Confirm */
3810 set_irn_pinned(store, op_pin_state_floats);
3812 if (get_Proj_proj(proj) == pn_Store_X_except) {
3813 DBG_OPT_EXC_REM(proj);
3814 return get_irg_bad(current_ir_graph);
3816 ir_node *blk = get_nodes_block(store);
3817 return new_r_Jmp(blk);
3823 } /* transform_node_Proj_Store */
3826 * Transform a Proj(Div) with a non-zero value.
3827 * Removes the exceptions and routes the memory to the NoMem node.
3829 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3830 ir_node *div = get_Proj_pred(proj);
3831 ir_node *b = get_Div_right(div);
3832 ir_node *res, *new_mem;
3833 const ir_node *confirm;
3836 if (value_not_zero(b, &confirm)) {
3837 /* div(x, y) && y != 0 */
3838 if (confirm == NULL) {
3839 /* we are sure we have a Const != 0 */
3840 new_mem = get_Div_mem(div);
3841 new_mem = skip_Pin(new_mem);
3842 set_Div_mem(div, new_mem);
3843 set_irn_pinned(div, op_pin_state_floats);
3846 proj_nr = get_Proj_proj(proj);
3848 case pn_Div_X_regular:
3849 return new_r_Jmp(get_nodes_block(div));
3851 case pn_Div_X_except:
3852 /* we found an exception handler, remove it */
3853 DBG_OPT_EXC_REM(proj);
3857 res = get_Div_mem(div);
3858 new_mem = get_irg_no_mem(current_ir_graph);
3861 /* This node can only float up to the Confirm block */
3862 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3864 set_irn_pinned(div, op_pin_state_floats);
3865 /* this is a Div without exception, we can remove the memory edge */
3866 set_Div_mem(div, new_mem);
3871 } /* transform_node_Proj_Div */
3874 * Transform a Proj(Mod) with a non-zero value.
3875 * Removes the exceptions and routes the memory to the NoMem node.
3877 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3878 ir_node *mod = get_Proj_pred(proj);
3879 ir_node *b = get_Mod_right(mod);
3880 ir_node *res, *new_mem;
3881 const ir_node *confirm;
3884 if (value_not_zero(b, &confirm)) {
3885 /* mod(x, y) && y != 0 */
3886 proj_nr = get_Proj_proj(proj);
3888 if (confirm == NULL) {
3889 /* we are sure we have a Const != 0 */
3890 new_mem = get_Mod_mem(mod);
3891 new_mem = skip_Pin(new_mem);
3892 set_Mod_mem(mod, new_mem);
3893 set_irn_pinned(mod, op_pin_state_floats);
3898 case pn_Mod_X_regular:
3899 return new_r_Jmp(get_irn_n(mod, -1));
3901 case pn_Mod_X_except:
3902 /* we found an exception handler, remove it */
3903 DBG_OPT_EXC_REM(proj);
3907 res = get_Mod_mem(mod);
3908 new_mem = get_irg_no_mem(current_ir_graph);
3911 /* This node can only float up to the Confirm block */
3912 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3914 /* this is a Mod without exception, we can remove the memory edge */
3915 set_Mod_mem(mod, new_mem);
3918 if (get_Mod_left(mod) == b) {
3919 /* a % a = 0 if a != 0 */
3920 ir_mode *mode = get_irn_mode(proj);
3921 ir_node *res = new_Const(get_mode_null(mode));
3923 DBG_OPT_CSTEVAL(mod, res);
3929 } /* transform_node_Proj_Mod */
3932 * Transform a Proj(DivMod) with a non-zero value.
3933 * Removes the exceptions and routes the memory to the NoMem node.
3935 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3936 ir_node *divmod = get_Proj_pred(proj);
3937 ir_node *b = get_DivMod_right(divmod);
3938 ir_node *res, *new_mem;
3939 const ir_node *confirm;
3942 if (value_not_zero(b, &confirm)) {
3943 /* DivMod(x, y) && y != 0 */
3944 proj_nr = get_Proj_proj(proj);
3946 if (confirm == NULL) {
3947 /* we are sure we have a Const != 0 */
3948 new_mem = get_DivMod_mem(divmod);
3949 new_mem = skip_Pin(new_mem);
3950 set_DivMod_mem(divmod, new_mem);
3951 set_irn_pinned(divmod, op_pin_state_floats);
3956 case pn_DivMod_X_regular:
3957 return new_r_Jmp(get_nodes_block(divmod));
3959 case pn_DivMod_X_except:
3960 /* we found an exception handler, remove it */
3961 DBG_OPT_EXC_REM(proj);
3965 res = get_DivMod_mem(divmod);
3966 new_mem = get_irg_no_mem(current_ir_graph);
3969 /* This node can only float up to the Confirm block */
3970 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3972 /* this is a DivMod without exception, we can remove the memory edge */
3973 set_DivMod_mem(divmod, new_mem);
3976 case pn_DivMod_res_mod:
3977 if (get_DivMod_left(divmod) == b) {
3978 /* a % a = 0 if a != 0 */
3979 ir_mode *mode = get_irn_mode(proj);
3980 ir_node *res = new_Const(get_mode_null(mode));
3982 DBG_OPT_CSTEVAL(divmod, res);
3988 } /* transform_node_Proj_DivMod */
3991 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3993 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3994 if (get_opt_unreachable_code()) {
3995 ir_node *n = get_Proj_pred(proj);
3996 ir_node *b = get_Cond_selector(n);
3998 if (mode_is_int(get_irn_mode(b))) {
3999 tarval *tb = value_of(b);
4001 if (tb != tarval_bad) {
4002 /* we have a constant switch */
4003 long num = get_Proj_proj(proj);
4005 if (num != get_Cond_default_proj(n)) { /* we cannot optimize default Proj's yet */
4006 if (get_tarval_long(tb) == num) {
4007 /* Do NOT create a jump here, or we will have 2 control flow ops
4008 * in a block. This case is optimized away in optimize_cf(). */
4011 /* this case will NEVER be taken, kill it */
4012 return get_irg_bad(current_ir_graph);
4016 long num = get_Proj_proj(proj);
4017 if (num != get_Cond_default_proj(n)) {
4018 /* Try handling with vrp data. We only remove dead parts. */
4019 tarval *tp = new_tarval_from_long(num, get_irn_mode(b));
4021 if (b->vrp.range_type == VRP_RANGE) {
4022 pn_Cmp cmp_result = tarval_cmp(b->vrp.range_bottom, tp);
4023 pn_Cmp cmp_result2 = tarval_cmp(b->vrp.range_top, tp);
4025 if ((cmp_result & pn_Cmp_Lt) == cmp_result && (cmp_result2
4026 & pn_Cmp_Gt) == cmp_result2) {
4027 return get_irg_bad(current_ir_graph);
4029 } else if (b->vrp.range_type == VRP_ANTIRANGE) {
4030 pn_Cmp cmp_result = tarval_cmp(b->vrp.range_bottom, tp);
4031 pn_Cmp cmp_result2 = tarval_cmp(b->vrp.range_top, tp);
4033 if ((cmp_result & pn_Cmp_Ge) == cmp_result && (cmp_result2
4034 & pn_Cmp_Le) == cmp_result2) {
4035 return get_irg_bad(current_ir_graph);
4040 tarval_and( b->vrp.bits_set, tp),
4044 return get_irg_bad(current_ir_graph);
4050 b->vrp.bits_not_set),
4051 b->vrp.bits_not_set)
4054 return get_irg_bad(current_ir_graph);
4063 } /* transform_node_Proj_Cond */
4066 * Create a 0 constant of given mode.
4068 static ir_node *create_zero_const(ir_mode *mode) {
4069 tarval *tv = get_mode_null(mode);
4070 ir_node *cnst = new_Const(tv);
4075 /* the order of the values is important! */
4076 typedef enum const_class {
4082 static const_class classify_const(const ir_node* n)
4084 if (is_Const(n)) return const_const;
4085 if (is_irn_constlike(n)) return const_like;
4090 * Determines whether r is more constlike or has a larger index (in that order)
4093 static int operands_are_normalized(const ir_node *l, const ir_node *r)
4095 const const_class l_order = classify_const(l);
4096 const const_class r_order = classify_const(r);
4098 l_order > r_order ||
4099 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
4103 * Normalizes and optimizes Cmp nodes.
4105 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
4106 ir_node *n = get_Proj_pred(proj);
4107 ir_node *left = get_Cmp_left(n);
4108 ir_node *right = get_Cmp_right(n);
4111 ir_mode *mode = NULL;
4112 long proj_nr = get_Proj_proj(proj);
4114 /* we can evaluate some cases directly */
4117 return new_Const(get_tarval_b_false());
4119 return new_Const(get_tarval_b_true());
4121 if (!mode_is_float(get_irn_mode(left)))
4122 return new_Const(get_tarval_b_true());
4128 /* remove Casts of both sides */
4129 left = skip_Cast(left);
4130 right = skip_Cast(right);
4132 /* Remove unnecessary conversions */
4133 /* TODO handle constants */
4134 if (is_Conv(left) && is_Conv(right)) {
4135 ir_mode *mode = get_irn_mode(left);
4136 ir_node *op_left = get_Conv_op(left);
4137 ir_node *op_right = get_Conv_op(right);
4138 ir_mode *mode_left = get_irn_mode(op_left);
4139 ir_mode *mode_right = get_irn_mode(op_right);
4141 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
4142 && mode_left != mode_b && mode_right != mode_b) {
4143 ir_node *block = get_nodes_block(n);
4145 if (mode_left == mode_right) {
4149 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
4150 } else if (smaller_mode(mode_left, mode_right)) {
4151 left = new_r_Conv(block, op_left, mode_right);
4154 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4155 } else if (smaller_mode(mode_right, mode_left)) {
4157 right = new_r_Conv(block, op_right, mode_left);
4159 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4164 /* remove operation on both sides if possible */
4165 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4167 * The following operations are NOT safe for floating point operations, for instance
4168 * 1.0 + inf == 2.0 + inf, =/=> x == y
4170 if (mode_is_int(get_irn_mode(left))) {
4171 unsigned lop = get_irn_opcode(left);
4173 if (lop == get_irn_opcode(right)) {
4174 ir_node *ll, *lr, *rl, *rr;
4176 /* same operation on both sides, try to remove */
4180 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
4181 left = get_unop_op(left);
4182 right = get_unop_op(right);
4184 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4187 ll = get_Add_left(left);
4188 lr = get_Add_right(left);
4189 rl = get_Add_left(right);
4190 rr = get_Add_right(right);
4193 /* X + a CMP X + b ==> a CMP b */
4197 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4198 } else if (ll == rr) {
4199 /* X + a CMP b + X ==> a CMP b */
4203 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4204 } else if (lr == rl) {
4205 /* a + X CMP X + b ==> a CMP b */
4209 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4210 } else if (lr == rr) {
4211 /* a + X CMP b + X ==> a CMP b */
4215 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4219 ll = get_Sub_left(left);
4220 lr = get_Sub_right(left);
4221 rl = get_Sub_left(right);
4222 rr = get_Sub_right(right);
4225 /* X - a CMP X - b ==> a CMP b */
4229 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4230 } else if (lr == rr) {
4231 /* a - X CMP b - X ==> a CMP b */
4235 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4239 if (get_Rotl_right(left) == get_Rotl_right(right)) {
4240 /* a ROTL X CMP b ROTL X ==> a CMP b */
4241 left = get_Rotl_left(left);
4242 right = get_Rotl_left(right);
4244 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4252 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4253 if (is_Add(left) || is_Sub(left)) {
4254 ir_node *ll = get_binop_left(left);
4255 ir_node *lr = get_binop_right(left);
4257 if (lr == right && is_Add(left)) {
4264 right = create_zero_const(get_irn_mode(left));
4266 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4269 if (is_Add(right) || is_Sub(right)) {
4270 ir_node *rl = get_binop_left(right);
4271 ir_node *rr = get_binop_right(right);
4273 if (rr == left && is_Add(right)) {
4280 right = create_zero_const(get_irn_mode(left));
4282 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4285 if (is_And(left) && is_Const(right)) {
4286 ir_node *ll = get_binop_left(left);
4287 ir_node *lr = get_binop_right(left);
4288 if (is_Shr(ll) && is_Const(lr)) {
4289 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
4290 ir_node *block = get_nodes_block(n);
4291 ir_mode *mode = get_irn_mode(left);
4293 ir_node *llr = get_Shr_right(ll);
4294 if (is_Const(llr)) {
4295 dbg_info *dbg = get_irn_dbg_info(left);
4297 tarval *c1 = get_Const_tarval(llr);
4298 tarval *c2 = get_Const_tarval(lr);
4299 tarval *c3 = get_Const_tarval(right);
4300 tarval *mask = tarval_shl(c2, c1);
4301 tarval *value = tarval_shl(c3, c1);
4303 left = new_rd_And(dbg, block, get_Shr_left(ll), new_Const(mask), mode);
4304 right = new_Const(value);
4309 } /* mode_is_int(...) */
4310 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
4312 /* replace mode_b compares with ands/ors */
4313 if (get_irn_mode(left) == mode_b) {
4314 ir_node *block = get_nodes_block(n);
4318 case pn_Cmp_Le: bres = new_r_Or( block, new_r_Not(block, left, mode_b), right, mode_b); break;
4319 case pn_Cmp_Lt: bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b); break;
4320 case pn_Cmp_Ge: bres = new_r_Or( block, left, new_r_Not(block, right, mode_b), mode_b); break;
4321 case pn_Cmp_Gt: bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b); break;
4322 case pn_Cmp_Lg: bres = new_r_Eor(block, left, right, mode_b); break;
4323 case pn_Cmp_Eq: bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b); break;
4324 default: bres = NULL;
4327 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4333 * First step: normalize the compare op
4334 * by placing the constant on the right side
4335 * or moving the lower address node to the left.
4337 if (!operands_are_normalized(left, right)) {
4343 proj_nr = get_inversed_pnc(proj_nr);
4348 * Second step: Try to reduce the magnitude
4349 * of a constant. This may help to generate better code
4350 * later and may help to normalize more compares.
4351 * Of course this is only possible for integer values.
4353 tv = value_of(right);
4354 if (tv != tarval_bad) {
4355 mode = get_irn_mode(right);
4357 /* TODO extend to arbitrary constants */
4358 if (is_Conv(left) && tarval_is_null(tv)) {
4359 ir_node *op = get_Conv_op(left);
4360 ir_mode *op_mode = get_irn_mode(op);
4363 * UpConv(x) REL 0 ==> x REL 0
4364 * Don't do this for float values as it's unclear whether it is a
4365 * win. (on the other side it makes detection/creation of fabs hard)
4367 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4368 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
4369 mode_is_signed(mode) || !mode_is_signed(op_mode)) &&
4370 !mode_is_float(mode)) {
4371 tv = get_mode_null(op_mode);
4375 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4379 if (tv != tarval_bad) {
4380 /* the following optimization is possible on modes without Overflow
4381 * on Unary Minus or on == and !=:
4382 * -a CMP c ==> a swap(CMP) -c
4384 * Beware: for two-complement Overflow may occur, so only == and != can
4385 * be optimized, see this:
4386 * -MININT < 0 =/=> MININT > 0 !!!
4388 if (is_Minus(left) &&
4389 (!mode_overflow_on_unary_Minus(mode) ||
4390 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
4391 tv = tarval_neg(tv);
4393 if (tv != tarval_bad) {
4394 left = get_Minus_op(left);
4395 proj_nr = get_inversed_pnc(proj_nr);
4397 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4399 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
4400 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4401 tv = tarval_not(tv);
4403 if (tv != tarval_bad) {
4404 left = get_Not_op(left);
4406 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4410 /* for integer modes, we have more */
4411 if (mode_is_int(mode)) {
4412 /* Ne includes Unordered which is not possible on integers.
4413 * However, frontends often use this wrong, so fix it here */
4414 if (proj_nr & pn_Cmp_Uo) {
4415 proj_nr &= ~pn_Cmp_Uo;
4416 set_Proj_proj(proj, proj_nr);
4419 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4420 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
4421 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
4422 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4424 if (tv != tarval_bad) {
4425 proj_nr ^= pn_Cmp_Eq;
4427 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4430 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4431 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
4432 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
4433 tv = tarval_add(tv, get_mode_one(mode));
4435 if (tv != tarval_bad) {
4436 proj_nr ^= pn_Cmp_Eq;
4438 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4442 /* the following reassociations work only for == and != */
4443 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4445 #if 0 /* Might be not that good in general */
4446 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4447 if (tarval_is_null(tv) && is_Sub(left)) {
4448 right = get_Sub_right(left);
4449 left = get_Sub_left(left);
4451 tv = value_of(right);
4453 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4457 if (tv != tarval_bad) {
4458 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4460 ir_node *c1 = get_Sub_right(left);
4461 tarval *tv2 = value_of(c1);
4463 if (tv2 != tarval_bad) {
4464 tv2 = tarval_add(tv, value_of(c1));
4466 if (tv2 != tarval_bad) {
4467 left = get_Sub_left(left);
4470 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4474 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4475 else if (is_Add(left)) {
4476 ir_node *a_l = get_Add_left(left);
4477 ir_node *a_r = get_Add_right(left);
4481 if (is_Const(a_l)) {
4483 tv2 = value_of(a_l);
4486 tv2 = value_of(a_r);
4489 if (tv2 != tarval_bad) {
4490 tv2 = tarval_sub(tv, tv2, NULL);
4492 if (tv2 != tarval_bad) {
4496 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4500 /* -a == c ==> a == -c, -a != c ==> a != -c */
4501 else if (is_Minus(left)) {
4502 tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4504 if (tv2 != tarval_bad) {
4505 left = get_Minus_op(left);
4508 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4513 /* the following reassociations work only for <= */
4514 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4515 if (tv != tarval_bad) {
4516 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4517 if (is_Abs(left)) { // TODO something is missing here
4523 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4524 switch (get_irn_opcode(left)) {
4528 c1 = get_And_right(left);
4531 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4532 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4534 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4536 /* TODO: move to constant evaluation */
4537 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4539 DBG_OPT_CSTEVAL(proj, c1);
4543 if (tarval_is_single_bit(tv)) {
4545 * optimization for AND:
4547 * And(x, C) == C ==> And(x, C) != 0
4548 * And(x, C) != C ==> And(X, C) == 0
4550 * if C is a single Bit constant.
4553 /* check for Constant's match. We have check hare the tarvals,
4554 because our const might be changed */
4555 if (get_Const_tarval(c1) == tv) {
4556 /* fine: do the transformation */
4557 tv = get_mode_null(get_tarval_mode(tv));
4558 proj_nr ^= pn_Cmp_Leg;
4560 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4566 c1 = get_Or_right(left);
4567 if (is_Const(c1) && tarval_is_null(tv)) {
4569 * Or(x, C) == 0 && C != 0 ==> FALSE
4570 * Or(x, C) != 0 && C != 0 ==> TRUE
4572 if (! tarval_is_null(get_Const_tarval(c1))) {
4573 /* TODO: move to constant evaluation */
4574 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4576 DBG_OPT_CSTEVAL(proj, c1);
4583 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4585 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4588 c1 = get_Shl_right(left);
4590 tarval *tv1 = get_Const_tarval(c1);
4591 ir_mode *mode = get_irn_mode(left);
4592 tarval *minus1 = get_mode_all_one(mode);
4593 tarval *amask = tarval_shr(minus1, tv1);
4594 tarval *cmask = tarval_shl(minus1, tv1);
4597 if (tarval_and(tv, cmask) != tv) {
4598 /* condition not met */
4599 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4601 DBG_OPT_CSTEVAL(proj, c1);
4604 sl = get_Shl_left(left);
4605 blk = get_nodes_block(n);
4606 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_Const(amask), mode);
4607 tv = tarval_shr(tv, tv1);
4609 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4614 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4616 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4619 c1 = get_Shr_right(left);
4621 tarval *tv1 = get_Const_tarval(c1);
4622 ir_mode *mode = get_irn_mode(left);
4623 tarval *minus1 = get_mode_all_one(mode);
4624 tarval *amask = tarval_shl(minus1, tv1);
4625 tarval *cmask = tarval_shr(minus1, tv1);
4628 if (tarval_and(tv, cmask) != tv) {
4629 /* condition not met */
4630 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4632 DBG_OPT_CSTEVAL(proj, c1);
4635 sl = get_Shr_left(left);
4636 blk = get_nodes_block(n);
4637 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_Const(amask), mode);
4638 tv = tarval_shl(tv, tv1);
4640 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4645 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4647 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4650 c1 = get_Shrs_right(left);
4652 tarval *tv1 = get_Const_tarval(c1);
4653 ir_mode *mode = get_irn_mode(left);
4654 tarval *minus1 = get_mode_all_one(mode);
4655 tarval *amask = tarval_shl(minus1, tv1);
4656 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4659 cond = tarval_sub(cond, tv1, NULL);
4660 cond = tarval_shrs(tv, cond);
4662 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4663 /* condition not met */
4664 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4666 DBG_OPT_CSTEVAL(proj, c1);
4669 sl = get_Shrs_left(left);
4670 blk = get_nodes_block(n);
4671 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_Const(amask), mode);
4672 tv = tarval_shl(tv, tv1);
4674 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4679 } /* tarval != bad */
4682 if (changed & 2) /* need a new Const */
4683 right = new_Const(tv);
4685 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4686 ir_node *op = get_Proj_pred(left);
4688 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4689 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4690 ir_node *c = get_binop_right(op);
4693 tarval *tv = get_Const_tarval(c);
4695 if (tarval_is_single_bit(tv)) {
4696 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4697 ir_node *v = get_binop_left(op);
4698 ir_node *blk = get_irn_n(op, -1);
4699 ir_mode *mode = get_irn_mode(v);
4701 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4702 left = new_rd_And(get_irn_dbg_info(op), blk, v, new_Const(tv), mode);
4704 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4711 ir_node *block = get_nodes_block(n);
4713 /* create a new compare */
4714 n = new_rd_Cmp(get_irn_dbg_info(n), block, left, right);
4715 proj = new_rd_Proj(get_irn_dbg_info(proj), block, n, get_irn_mode(proj), proj_nr);
4719 } /* transform_node_Proj_Cmp */
4722 * Optimize CopyB(mem, x, x) into a Nop.
4724 static ir_node *transform_node_Proj_CopyB(ir_node *proj) {
4725 ir_node *copyb = get_Proj_pred(proj);
4726 ir_node *a = get_CopyB_dst(copyb);
4727 ir_node *b = get_CopyB_src(copyb);
4730 switch (get_Proj_proj(proj)) {
4731 case pn_CopyB_X_regular:
4732 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4733 DBG_OPT_EXC_REM(proj);
4734 proj = new_r_Jmp(get_nodes_block(copyb));
4736 case pn_CopyB_X_except:
4737 DBG_OPT_EXC_REM(proj);
4738 proj = get_irg_bad(get_irn_irg(proj));
4745 } /* transform_node_Proj_CopyB */
4748 * Optimize Bounds(idx, idx, upper) into idx.
4750 static ir_node *transform_node_Proj_Bound(ir_node *proj) {
4751 ir_node *oldn = proj;
4752 ir_node *bound = get_Proj_pred(proj);
4753 ir_node *idx = get_Bound_index(bound);
4754 ir_node *pred = skip_Proj(idx);
4757 if (idx == get_Bound_lower(bound))
4759 else if (is_Bound(pred)) {
4761 * idx was Bounds checked in the same MacroBlock previously,
4762 * it is still valid if lower <= pred_lower && pred_upper <= upper.
4764 ir_node *lower = get_Bound_lower(bound);
4765 ir_node *upper = get_Bound_upper(bound);
4766 if (get_Bound_lower(pred) == lower &&
4767 get_Bound_upper(pred) == upper &&
4768 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
4770 * One could expect that we simply return the previous
4771 * Bound here. However, this would be wrong, as we could
4772 * add an exception Proj to a new location then.
4773 * So, we must turn in into a tuple.
4779 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4780 switch (get_Proj_proj(proj)) {
4782 DBG_OPT_EXC_REM(proj);
4783 proj = get_Bound_mem(bound);
4785 case pn_Bound_X_except:
4786 DBG_OPT_EXC_REM(proj);
4787 proj = get_irg_bad(get_irn_irg(proj));
4791 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4793 case pn_Bound_X_regular:
4794 DBG_OPT_EXC_REM(proj);
4795 proj = new_r_Jmp(get_nodes_block(bound));
4802 } /* transform_node_Proj_Bound */
4805 * Does all optimizations on nodes that must be done on it's Proj's
4806 * because of creating new nodes.
4808 static ir_node *transform_node_Proj(ir_node *proj) {
4809 ir_node *n = get_Proj_pred(proj);
4811 if (n->op->ops.transform_node_Proj)
4812 return n->op->ops.transform_node_Proj(proj);
4814 } /* transform_node_Proj */
4817 * Move Confirms down through Phi nodes.
4819 static ir_node *transform_node_Phi(ir_node *phi) {
4821 ir_mode *mode = get_irn_mode(phi);
4823 if (mode_is_reference(mode)) {
4824 n = get_irn_arity(phi);
4826 /* Beware of Phi0 */
4828 ir_node *pred = get_irn_n(phi, 0);
4829 ir_node *bound, *new_Phi, *block, **in;
4832 if (! is_Confirm(pred))
4835 bound = get_Confirm_bound(pred);
4836 pnc = get_Confirm_cmp(pred);
4838 NEW_ARR_A(ir_node *, in, n);
4839 in[0] = get_Confirm_value(pred);
4841 for (i = 1; i < n; ++i) {
4842 pred = get_irn_n(phi, i);
4844 if (! is_Confirm(pred) ||
4845 get_Confirm_bound(pred) != bound ||
4846 get_Confirm_cmp(pred) != pnc)
4848 in[i] = get_Confirm_value(pred);
4850 /* move the Confirm nodes "behind" the Phi */
4851 block = get_irn_n(phi, -1);
4852 new_Phi = new_r_Phi(block, n, in, get_irn_mode(phi));
4853 return new_r_Confirm(block, new_Phi, bound, pnc);
4857 } /* transform_node_Phi */
4860 * Returns the operands of a commutative bin-op, if one operand is
4861 * a const, it is returned as the second one.
4863 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4864 ir_node *op_a = get_binop_left(binop);
4865 ir_node *op_b = get_binop_right(binop);
4867 assert(is_op_commutative(get_irn_op(binop)));
4869 if (is_Const(op_a)) {
4876 } /* get_comm_Binop_Ops */
4879 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4880 * Such pattern may arise in bitfield stores.
4882 * value c4 value c4 & c2
4883 * AND c3 AND c1 | c3
4890 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4893 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4896 ir_node *and_l, *c3;
4897 ir_node *value, *c4;
4898 ir_node *new_and, *new_const, *block;
4899 ir_mode *mode = get_irn_mode(or);
4901 tarval *tv1, *tv2, *tv3, *tv4, *tv;
4904 get_comm_Binop_Ops(or, &and, &c1);
4905 if (!is_Const(c1) || !is_And(and))
4908 get_comm_Binop_Ops(and, &or_l, &c2);
4912 tv1 = get_Const_tarval(c1);
4913 tv2 = get_Const_tarval(c2);
4915 tv = tarval_or(tv1, tv2);
4916 if (tarval_is_all_one(tv)) {
4917 /* the AND does NOT clear a bit with isn't set by the OR */
4918 set_Or_left(or, or_l);
4919 set_Or_right(or, c1);
4921 /* check for more */
4928 get_comm_Binop_Ops(or_l, &and_l, &c3);
4929 if (!is_Const(c3) || !is_And(and_l))
4932 get_comm_Binop_Ops(and_l, &value, &c4);
4936 /* ok, found the pattern, check for conditions */
4937 assert(mode == get_irn_mode(and));
4938 assert(mode == get_irn_mode(or_l));
4939 assert(mode == get_irn_mode(and_l));
4941 tv3 = get_Const_tarval(c3);
4942 tv4 = get_Const_tarval(c4);
4944 tv = tarval_or(tv4, tv2);
4945 if (!tarval_is_all_one(tv)) {
4946 /* have at least one 0 at the same bit position */
4950 if (tv3 != tarval_andnot(tv3, tv4)) {
4951 /* bit in the or_mask is outside the and_mask */
4955 if (tv1 != tarval_andnot(tv1, tv2)) {
4956 /* bit in the or_mask is outside the and_mask */
4960 /* ok, all conditions met */
4961 block = get_irn_n(or, -1);
4963 new_and = new_r_And(block, value, new_Const(tarval_and(tv4, tv2)), mode);
4965 new_const = new_Const(tarval_or(tv3, tv1));
4967 set_Or_left(or, new_and);
4968 set_Or_right(or, new_const);
4970 /* check for more */
4972 } /* transform_node_Or_bf_store */
4975 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4977 static ir_node *transform_node_Or_Rotl(ir_node *or) {
4978 ir_mode *mode = get_irn_mode(or);
4979 ir_node *shl, *shr, *block;
4980 ir_node *irn, *x, *c1, *c2, *v, *sub, *n, *rotval;
4983 if (! mode_is_int(mode))
4986 shl = get_binop_left(or);
4987 shr = get_binop_right(or);
4996 } else if (!is_Shl(shl)) {
4998 } else if (!is_Shr(shr)) {
5001 x = get_Shl_left(shl);
5002 if (x != get_Shr_left(shr))
5005 c1 = get_Shl_right(shl);
5006 c2 = get_Shr_right(shr);
5007 if (is_Const(c1) && is_Const(c2)) {
5008 tv1 = get_Const_tarval(c1);
5009 if (! tarval_is_long(tv1))
5012 tv2 = get_Const_tarval(c2);
5013 if (! tarval_is_long(tv2))
5016 if (get_tarval_long(tv1) + get_tarval_long(tv2)
5017 != (int) get_mode_size_bits(mode))
5020 /* yet, condition met */
5021 block = get_nodes_block(or);
5023 n = new_r_Rotl(block, x, c1, mode);
5025 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
5032 rotval = sub; /* a Rot right is not supported, so use a rot left */
5033 } else if (is_Sub(c2)) {
5039 if (get_Sub_right(sub) != v)
5042 c1 = get_Sub_left(sub);
5046 tv1 = get_Const_tarval(c1);
5047 if (! tarval_is_long(tv1))
5050 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
5053 /* yet, condition met */
5054 block = get_nodes_block(or);
5056 n = new_r_Rotl(block, x, rotval, mode);
5058 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL);
5060 } /* transform_node_Or_Rotl */
5065 static ir_node *transform_node_Or(ir_node *n) {
5066 ir_node *c, *oldn = n;
5067 ir_node *a = get_Or_left(n);
5068 ir_node *b = get_Or_right(n);
5071 if (is_Not(a) && is_Not(b)) {
5072 /* ~a | ~b = ~(a&b) */
5073 ir_node *block = get_nodes_block(n);
5075 mode = get_irn_mode(n);
5078 n = new_rd_And(get_irn_dbg_info(n), block, a, b, mode);
5079 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
5080 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
5084 /* we can evaluate 2 Projs of the same Cmp */
5085 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
5086 ir_node *pred_a = get_Proj_pred(a);
5087 ir_node *pred_b = get_Proj_pred(b);
5088 if (pred_a == pred_b) {
5089 dbg_info *dbgi = get_irn_dbg_info(n);
5090 ir_node *block = get_nodes_block(pred_a);
5091 pn_Cmp pn_a = get_Proj_proj(a);
5092 pn_Cmp pn_b = get_Proj_proj(b);
5093 /* yes, we can simply calculate with pncs */
5094 pn_Cmp new_pnc = pn_a | pn_b;
5096 return new_rd_Proj(dbgi, block, pred_a, mode_b, new_pnc);
5100 mode = get_irn_mode(n);
5101 HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode);
5103 n = transform_node_Or_bf_store(n);
5104 n = transform_node_Or_Rotl(n);
5108 n = transform_bitwise_distributive(n, transform_node_Or);
5111 } /* transform_node_Or */
5115 static ir_node *transform_node(ir_node *n);
5118 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
5120 * Should be moved to reassociation?
5122 static ir_node *transform_node_shift(ir_node *n) {
5123 ir_node *left, *right;
5125 tarval *tv1, *tv2, *res;
5126 ir_node *in[2], *irn, *block;
5128 left = get_binop_left(n);
5130 /* different operations */
5131 if (get_irn_op(left) != get_irn_op(n))
5134 right = get_binop_right(n);
5135 tv1 = value_of(right);
5136 if (tv1 == tarval_bad)
5139 tv2 = value_of(get_binop_right(left));
5140 if (tv2 == tarval_bad)
5143 res = tarval_add(tv1, tv2);
5144 mode = get_irn_mode(n);
5146 /* beware: a simple replacement works only, if res < modulo shift */
5148 int modulo_shf = get_mode_modulo_shift(mode);
5149 if (modulo_shf > 0) {
5150 tarval *modulo = new_tarval_from_long(modulo_shf,
5151 get_tarval_mode(res));
5153 assert(modulo_shf >= (int) get_mode_size_bits(mode));
5155 /* shifting too much */
5156 if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) {
5158 ir_node *block = get_nodes_block(n);
5159 dbg_info *dbgi = get_irn_dbg_info(n);
5160 ir_mode *smode = get_irn_mode(right);
5161 ir_node *cnst = new_Const_long(smode, get_mode_size_bits(mode) - 1);
5162 return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode);
5165 return new_Const(get_mode_null(mode));
5169 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
5172 /* ok, we can replace it */
5173 block = get_nodes_block(n);
5175 in[0] = get_binop_left(left);
5176 in[1] = new_Const(res);
5178 irn = new_ir_node(NULL, get_Block_irg(block), block, get_irn_op(n), mode, 2, in);
5180 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5182 return transform_node(irn);
5183 } /* transform_node_shift */
5186 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
5188 * - and, or, xor instead of &
5189 * - Shl, Shr, Shrs, rotl instead of >>
5190 * (with a special case for Or/Xor + Shrs)
5192 static ir_node *transform_node_bitop_shift(ir_node *n) {
5194 ir_node *right = get_binop_right(n);
5195 ir_mode *mode = get_irn_mode(n);
5196 ir_node *bitop_left;
5197 ir_node *bitop_right;
5208 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
5210 if (!is_Const(right))
5213 left = get_binop_left(n);
5214 op_left = get_irn_op(left);
5215 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
5218 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
5219 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
5220 /* TODO: test if sign bit is affectes */
5224 bitop_right = get_binop_right(left);
5225 if (!is_Const(bitop_right))
5228 bitop_left = get_binop_left(left);
5230 block = get_nodes_block(n);
5231 dbgi = get_irn_dbg_info(n);
5232 tv1 = get_Const_tarval(bitop_right);
5233 tv2 = get_Const_tarval(right);
5235 assert(get_tarval_mode(tv1) == mode);
5238 new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode);
5239 tv_shift = tarval_shl(tv1, tv2);
5240 } else if(is_Shr(n)) {
5241 new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode);
5242 tv_shift = tarval_shr(tv1, tv2);
5243 } else if(is_Shrs(n)) {
5244 new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode);
5245 tv_shift = tarval_shrs(tv1, tv2);
5248 new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode);
5249 tv_shift = tarval_rotl(tv1, tv2);
5252 assert(get_tarval_mode(tv_shift) == mode);
5253 new_const = new_Const(tv_shift);
5255 if (op_left == op_And) {
5256 new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode);
5257 } else if(op_left == op_Or) {
5258 new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode);
5260 assert(op_left == op_Eor);
5261 new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode);
5269 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5271 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5272 * (also with x >>s c1 when c1>=c2)
5274 static ir_node *transform_node_shl_shr(ir_node *n) {
5276 ir_node *right = get_binop_right(n);
5292 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5294 if (!is_Const(right))
5297 left = get_binop_left(n);
5298 mode = get_irn_mode(n);
5299 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5300 ir_node *shr_right = get_binop_right(left);
5302 if (!is_Const(shr_right))
5305 x = get_binop_left(left);
5306 tv_shr = get_Const_tarval(shr_right);
5307 tv_shl = get_Const_tarval(right);
5309 if (is_Shrs(left)) {
5310 /* shrs variant only allowed if c1 >= c2 */
5311 if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge))
5314 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5317 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5319 tv_mask = tarval_shl(tv_mask, tv_shl);
5320 } else if(is_Shr(n) && is_Shl(left)) {
5321 ir_node *shl_right = get_Shl_right(left);
5323 if (!is_Const(shl_right))
5326 x = get_Shl_left(left);
5327 tv_shr = get_Const_tarval(right);
5328 tv_shl = get_Const_tarval(shl_right);
5330 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5331 tv_mask = tarval_shr(tv_mask, tv_shr);
5336 if (get_tarval_mode(tv_shl) != get_tarval_mode(tv_shr)) {
5337 tv_shl = tarval_convert_to(tv_shl, get_tarval_mode(tv_shr));
5340 assert(tv_mask != tarval_bad);
5341 assert(get_tarval_mode(tv_mask) == mode);
5343 irg = get_irn_irg(n);
5344 block = get_nodes_block(n);
5345 dbgi = get_irn_dbg_info(n);
5347 pnc = tarval_cmp(tv_shl, tv_shr);
5348 if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) {
5349 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5350 new_const = new_Const(tv_shift);
5352 new_shift = new_rd_Shrs(dbgi, block, x, new_const, mode);
5354 new_shift = new_rd_Shr(dbgi, block, x, new_const, mode);
5357 assert(pnc == pn_Cmp_Gt);
5358 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5359 new_const = new_Const(tv_shift);
5360 new_shift = new_rd_Shl(dbgi, block, x, new_const, mode);
5363 new_const = new_Const(tv_mask);
5364 new_and = new_rd_And(dbgi, block, new_shift, new_const, mode);
5372 static ir_node *transform_node_Shr(ir_node *n) {
5373 ir_node *c, *oldn = n;
5374 ir_node *left = get_Shr_left(n);
5375 ir_node *right = get_Shr_right(n);
5376 ir_mode *mode = get_irn_mode(n);
5378 HANDLE_BINOP_PHI((eval_func) tarval_shr, left, right, c, mode);
5379 n = transform_node_shift(n);
5382 n = transform_node_shl_shr(n);
5384 n = transform_node_bitop_shift(n);
5387 } /* transform_node_Shr */
5392 static ir_node *transform_node_Shrs(ir_node *n) {
5393 ir_node *c, *oldn = n;
5394 ir_node *a = get_Shrs_left(n);
5395 ir_node *b = get_Shrs_right(n);
5396 ir_mode *mode = get_irn_mode(n);
5398 HANDLE_BINOP_PHI((eval_func) tarval_shrs, a, b, c, mode);
5399 n = transform_node_shift(n);
5402 n = transform_node_bitop_shift(n);
5405 } /* transform_node_Shrs */
5410 static ir_node *transform_node_Shl(ir_node *n) {
5411 ir_node *c, *oldn = n;
5412 ir_node *a = get_Shl_left(n);
5413 ir_node *b = get_Shl_right(n);
5414 ir_mode *mode = get_irn_mode(n);
5416 HANDLE_BINOP_PHI((eval_func) tarval_shl, a, b, c, mode);
5417 n = transform_node_shift(n);
5420 n = transform_node_shl_shr(n);
5422 n = transform_node_bitop_shift(n);
5425 } /* transform_node_Shl */
5430 static ir_node *transform_node_Rotl(ir_node *n) {
5431 ir_node *c, *oldn = n;
5432 ir_node *a = get_Rotl_left(n);
5433 ir_node *b = get_Rotl_right(n);
5434 ir_mode *mode = get_irn_mode(n);
5436 HANDLE_BINOP_PHI((eval_func) tarval_rotl, a, b, c, mode);
5437 n = transform_node_shift(n);
5440 n = transform_node_bitop_shift(n);
5443 } /* transform_node_Rotl */
5448 static ir_node *transform_node_Conv(ir_node *n) {
5449 ir_node *c, *oldn = n;
5450 ir_mode *mode = get_irn_mode(n);
5451 ir_node *a = get_Conv_op(n);
5453 if (mode != mode_b && is_const_Phi(a)) {
5454 /* Do NOT optimize mode_b Conv's, this leads to remaining
5455 * Phib nodes later, because the conv_b_lower operation
5456 * is instantly reverted, when it tries to insert a Convb.
5458 c = apply_conv_on_phi(a, mode);
5460 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5465 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5466 return new_r_Unknown(current_ir_graph, mode);
5469 if (mode_is_reference(mode) &&
5470 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5472 ir_node *l = get_Add_left(a);
5473 ir_node *r = get_Add_right(a);
5474 dbg_info *dbgi = get_irn_dbg_info(a);
5475 ir_node *block = get_nodes_block(n);
5477 ir_node *lop = get_Conv_op(l);
5478 if(get_irn_mode(lop) == mode) {
5479 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5480 n = new_rd_Add(dbgi, block, lop, r, mode);
5485 ir_node *rop = get_Conv_op(r);
5486 if(get_irn_mode(rop) == mode) {
5487 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5488 n = new_rd_Add(dbgi, block, l, rop, mode);
5495 } /* transform_node_Conv */
5498 * Remove dead blocks and nodes in dead blocks
5499 * in keep alive list. We do not generate a new End node.
5501 static ir_node *transform_node_End(ir_node *n) {
5502 int i, j, n_keepalives = get_End_n_keepalives(n);
5505 NEW_ARR_A(ir_node *, in, n_keepalives);
5507 for (i = j = 0; i < n_keepalives; ++i) {
5508 ir_node *ka = get_End_keepalive(n, i);
5510 if (! is_Block_dead(ka)) {
5514 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
5516 } else if (is_Bad(ka)) {
5517 /* no need to keep Bad */
5522 if (j != n_keepalives)
5523 set_End_keepalives(n, j, in);
5525 } /* transform_node_End */
5527 bool is_negated_value(ir_node *a, ir_node *b)
5529 if (is_Minus(a) && get_Minus_op(a) == b)
5531 if (is_Minus(b) && get_Minus_op(b) == a)
5533 if (is_Sub(a) && is_Sub(b)) {
5534 ir_node *a_left = get_Sub_left(a);
5535 ir_node *a_right = get_Sub_right(a);
5536 ir_node *b_left = get_Sub_left(b);
5537 ir_node *b_right = get_Sub_right(b);
5539 if (a_left == b_right && a_right == b_left)
5547 * Optimize a Mux into some simpler cases.
5549 static ir_node *transform_node_Mux(ir_node *n) {
5550 ir_node *oldn = n, *sel = get_Mux_sel(n);
5551 ir_mode *mode = get_irn_mode(n);
5552 ir_node *t = get_Mux_true(n);
5553 ir_node *f = get_Mux_false(n);
5554 ir_graph *irg = current_ir_graph;
5556 if (is_irg_state(irg, IR_GRAPH_STATE_KEEP_MUX))
5560 ir_node* block = get_nodes_block(n);
5562 ir_node* c1 = get_Mux_sel(t);
5563 ir_node* t1 = get_Mux_true(t);
5564 ir_node* f1 = get_Mux_false(t);
5566 /* Mux(cond0, Mux(cond1, x, y), y) -> typical if (cond0 && cond1) x else y */
5567 ir_node* and_ = new_r_And(block, c0, c1, mode_b);
5568 ir_node* new_mux = new_r_Mux(block, and_, f1, t1, mode);
5573 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5574 } else if (f == t1) {
5575 /* Mux(cond0, Mux(cond1, x, y), x) */
5576 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5577 ir_node* and_ = new_r_And(block, c0, not_c1, mode_b);
5578 ir_node* new_mux = new_r_Mux(block, and_, t1, f1, mode);
5583 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5585 } else if (is_Mux(f)) {
5586 ir_node* block = get_nodes_block(n);
5588 ir_node* c1 = get_Mux_sel(f);
5589 ir_node* t1 = get_Mux_true(f);
5590 ir_node* f1 = get_Mux_false(f);
5592 /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */
5593 ir_node* or_ = new_r_Or(block, c0, c1, mode_b);
5594 ir_node* new_mux = new_r_Mux(block, or_, f1, t1, mode);
5599 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5600 } else if (t == f1) {
5601 /* Mux(cond0, x, Mux(cond1, y, x)) */
5602 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5603 ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b);
5604 ir_node* new_mux = new_r_Mux(block, or_, t1, f1, mode);
5609 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5613 /* first normalization step: move a possible zero to the false case */
5615 ir_node *cmp = get_Proj_pred(sel);
5618 if (is_Const(t) && is_Const_null(t)) {
5621 /* Mux(x, 0, y) => Mux(x, y, 0) */
5622 pn_Cmp pnc = get_Proj_proj(sel);
5623 sel = new_r_Proj(get_nodes_block(cmp), cmp, mode_b,
5624 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
5625 n = new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, t, f, mode);
5633 /* note: after normalization, false can only happen on default */
5634 if (mode == mode_b) {
5635 dbg_info *dbg = get_irn_dbg_info(n);
5636 ir_node *block = get_nodes_block(n);
5639 tarval *tv_t = get_Const_tarval(t);
5640 if (tv_t == tarval_b_true) {
5642 /* Muxb(sel, true, false) = sel */
5643 assert(get_Const_tarval(f) == tarval_b_false);
5644 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5647 /* Muxb(sel, true, x) = Or(sel, x) */
5648 n = new_rd_Or(dbg, block, sel, f, mode_b);
5649 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5653 } else if (is_Const(f)) {
5654 tarval *tv_f = get_Const_tarval(f);
5655 if (tv_f == tarval_b_true) {
5656 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5657 ir_node* not_sel = new_rd_Not(dbg, block, sel, mode_b);
5658 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5659 n = new_rd_Or(dbg, block, not_sel, t, mode_b);
5662 /* Muxb(sel, x, false) = And(sel, x) */
5663 assert(tv_f == tarval_b_false);
5664 n = new_rd_And(dbg, block, sel, t, mode_b);
5665 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5671 /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */
5672 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5673 tarval *a = get_Const_tarval(t);
5674 tarval *b = get_Const_tarval(f);
5677 if (tarval_is_one(a) && tarval_is_null(b)) {
5678 ir_node *block = get_nodes_block(n);
5679 ir_node *conv = new_r_Conv(block, sel, mode);
5681 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5683 } else if (tarval_is_null(a) && tarval_is_one(b)) {
5684 ir_node *block = get_nodes_block(n);
5685 ir_node *not_ = new_r_Not(block, sel, mode_b);
5686 ir_node *conv = new_r_Conv(block, not_, mode);
5688 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5691 /* TODO: it's not really clear if that helps in general or should be moved
5692 * to backend, especially with the MUX->Conv transformation above */
5693 if (tarval_cmp(a, b) & pn_Cmp_Gt) {
5694 diff = tarval_sub(a, b, NULL);
5697 diff = tarval_sub(b, a, NULL);
5701 if (diff == get_tarval_one(mode)) {
5702 dbg_info *dbg = get_irn_dbg_info(n);
5703 ir_node *block = get_nodes_block(n);
5704 ir_node *t = new_Const(tarval_sub(a, min, NULL));
5705 ir_node *f = new_Const(tarval_sub(b, min, NULL));
5706 n = new_rd_Mux(dbg, block, sel, f, t, mode);
5707 n = new_rd_Add(dbg, block, n, new_Const(min), mode);
5713 ir_node *cmp = get_Proj_pred(sel);
5714 long pn = get_Proj_proj(sel);
5717 * Note: normalization puts the constant on the right side,
5718 * so we check only one case.
5720 * Note further that these optimization work even for floating point
5721 * with NaN's because -NaN == NaN.
5722 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5726 ir_node *cmp_r = get_Cmp_right(cmp);
5727 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5728 ir_node *block = get_nodes_block(n);
5729 ir_node *cmp_l = get_Cmp_left(cmp);
5731 if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
5734 /* NaN's work fine with abs, so it is ok to remove Uo */
5735 long pnc = pn & ~pn_Cmp_Uo;
5737 if ( (cmp_l == t && (pnc == pn_Cmp_Ge || pnc == pn_Cmp_Gt))
5738 || (cmp_l == f && (pnc == pn_Cmp_Le || pnc == pn_Cmp_Lt)))
5740 /* Mux(a >/>= 0, a, -a) = Mux(a </<= 0, -a, a) ==> Abs(a) */
5741 n = new_rd_Abs(get_irn_dbg_info(n), block, cmp_l, mode);
5742 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5744 } else if ((cmp_l == t && (pnc == pn_Cmp_Le || pnc == pn_Cmp_Lt))
5745 || (cmp_l == f && (pnc == pn_Cmp_Ge || pnc == pn_Cmp_Gt)))
5747 /* Mux(a </<= 0, a, -a) = Mux(a >/>= 0, -a, a) ==> -Abs(a) */
5748 n = new_rd_Abs(get_irn_dbg_info(n), block, cmp_l, mode);
5749 n = new_rd_Minus(get_irn_dbg_info(n), block, n, mode);
5750 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5755 if (mode_is_int(mode)) {
5757 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
5758 /* Mux((a & b) != 0, c, 0) */
5759 ir_node *and_r = get_And_right(cmp_l);
5762 if (and_r == t && f == cmp_r) {
5763 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5764 if (pn == pn_Cmp_Lg) {
5765 /* Mux((a & 2^C) != 0, 2^C, 0) */
5767 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5769 /* Mux((a & 2^C) == 0, 2^C, 0) */
5770 n = new_rd_Eor(get_irn_dbg_info(n),
5771 block, cmp_l, t, mode);
5772 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5777 if (is_Shl(and_r)) {
5778 ir_node *shl_l = get_Shl_left(and_r);
5779 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5780 if (and_r == t && f == cmp_r) {
5781 if (pn == pn_Cmp_Lg) {
5782 /* (a & (1 << n)) != 0, (1 << n), 0) */
5784 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5786 /* (a & (1 << n)) == 0, (1 << n), 0) */
5787 n = new_rd_Eor(get_irn_dbg_info(n),
5788 block, cmp_l, t, mode);
5789 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5795 and_l = get_And_left(cmp_l);
5796 if (is_Shl(and_l)) {
5797 ir_node *shl_l = get_Shl_left(and_l);
5798 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5799 if (and_l == t && f == cmp_r) {
5800 if (pn == pn_Cmp_Lg) {
5801 /* ((1 << n) & a) != 0, (1 << n), 0) */
5803 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5805 /* ((1 << n) & a) == 0, (1 << n), 0) */
5806 n = new_rd_Eor(get_irn_dbg_info(n),
5807 block, cmp_l, t, mode);
5808 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5821 } /* transform_node_Mux */
5824 * optimize Sync nodes that have other syncs as input we simply add the inputs
5825 * of the other sync to our own inputs
5827 static ir_node *transform_node_Sync(ir_node *n) {
5828 int arity = get_Sync_n_preds(n);
5831 for (i = 0; i < arity;) {
5832 ir_node *pred = get_Sync_pred(n, i);
5836 if (!is_Sync(pred)) {
5844 pred_arity = get_Sync_n_preds(pred);
5845 for (j = 0; j < pred_arity; ++j) {
5846 ir_node *pred_pred = get_Sync_pred(pred, j);
5851 add_irn_n(n, pred_pred);
5855 if (get_Sync_pred(n, k) == pred_pred) break;
5860 /* rehash the sync node */
5861 add_identities(current_ir_graph->value_table, n);
5864 } /* transform_node_Sync */
5867 * optimize a trampoline Call into a direct Call
5869 static ir_node *transform_node_Call(ir_node *call) {
5870 ir_node *callee = get_Call_ptr(call);
5871 ir_node *adr, *mem, *res, *bl, **in;
5872 ir_type *ctp, *mtp, *tp;
5875 int i, n_res, n_param;
5878 if (! is_Proj(callee))
5880 callee = get_Proj_pred(callee);
5881 if (! is_Builtin(callee))
5883 if (get_Builtin_kind(callee) != ir_bk_inner_trampoline)
5886 mem = get_Call_mem(call);
5888 if (skip_Proj(mem) == callee) {
5889 /* memory is routed to the trampoline, skip */
5890 mem = get_Builtin_mem(callee);
5893 /* build a new call type */
5894 mtp = get_Call_type(call);
5895 tdb = get_type_dbg_info(mtp);
5897 n_res = get_method_n_ress(mtp);
5898 n_param = get_method_n_params(mtp);
5899 ctp = new_d_type_method(n_param + 1, n_res, tdb);
5901 for (i = 0; i < n_res; ++i)
5902 set_method_res_type(ctp, i, get_method_res_type(mtp, i));
5904 NEW_ARR_A(ir_node *, in, n_param + 1);
5906 /* FIXME: we don't need a new pointer type in every step */
5907 tp = get_irg_frame_type(current_ir_graph);
5908 tp = new_type_pointer(tp);
5909 set_method_param_type(ctp, 0, tp);
5911 in[0] = get_Builtin_param(callee, 2);
5912 for (i = 0; i < n_param; ++i) {
5913 set_method_param_type(ctp, i + 1, get_method_param_type(mtp, i));
5914 in[i + 1] = get_Call_param(call, i);
5916 var = get_method_variadicity(mtp);
5917 set_method_variadicity(ctp, var);
5918 if (var == variadicity_variadic) {
5919 set_method_first_variadic_param_index(ctp, get_method_first_variadic_param_index(mtp) + 1);
5921 /* When we resolve a trampoline, the function must be called by a this-call */
5922 set_method_calling_convention(ctp, get_method_calling_convention(mtp) | cc_this_call);
5923 set_method_additional_properties(ctp, get_method_additional_properties(mtp));
5925 adr = get_Builtin_param(callee, 1);
5927 db = get_irn_dbg_info(call);
5928 bl = get_nodes_block(call);
5930 res = new_rd_Call(db, bl, mem, adr, n_param + 1, in, ctp);
5931 if (get_irn_pinned(call) == op_pin_state_floats)
5932 set_irn_pinned(res, op_pin_state_floats);
5934 } /* transform_node_Call */
5937 * Tries several [inplace] [optimizing] transformations and returns an
5938 * equivalent node. The difference to equivalent_node() is that these
5939 * transformations _do_ generate new nodes, and thus the old node must
5940 * not be freed even if the equivalent node isn't the old one.
5942 static ir_node *transform_node(ir_node *n) {
5946 * Transform_node is the only "optimizing transformation" that might
5947 * return a node with a different opcode. We iterate HERE until fixpoint
5948 * to get the final result.
5952 if (n->op->ops.transform_node != NULL)
5953 n = n->op->ops.transform_node(n);
5954 } while (oldn != n);
5957 } /* transform_node */
5960 * Sets the default transform node operation for an ir_op_ops.
5962 * @param code the opcode for the default operation
5963 * @param ops the operations initialized
5968 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5972 ops->transform_node = transform_node_##a; \
5974 #define CASE_PROJ(a) \
5976 ops->transform_node_Proj = transform_node_Proj_##a; \
5978 #define CASE_PROJ_EX(a) \
5980 ops->transform_node = transform_node_##a; \
5981 ops->transform_node_Proj = transform_node_Proj_##a; \
5990 CASE_PROJ_EX(DivMod);
6025 } /* firm_set_default_transform_node */
6028 /* **************** Common Subexpression Elimination **************** */
6030 /** The size of the hash table used, should estimate the number of nodes
6032 #define N_IR_NODES 512
6034 /** Compares the attributes of two Const nodes. */
6035 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
6036 return (get_Const_tarval(a) != get_Const_tarval(b))
6037 || (get_Const_type(a) != get_Const_type(b));
6038 } /* node_cmp_attr_Const */
6040 /** Compares the attributes of two Proj nodes. */
6041 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
6042 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
6043 } /* node_cmp_attr_Proj */
6045 /** Compares the attributes of two Filter nodes. */
6046 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
6047 return get_Filter_proj(a) != get_Filter_proj(b);
6048 } /* node_cmp_attr_Filter */
6050 /** Compares the attributes of two Alloc nodes. */
6051 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
6052 const alloc_attr *pa = get_irn_alloc_attr(a);
6053 const alloc_attr *pb = get_irn_alloc_attr(b);
6054 return (pa->where != pb->where) || (pa->type != pb->type);
6055 } /* node_cmp_attr_Alloc */
6057 /** Compares the attributes of two Free nodes. */
6058 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
6059 const free_attr *pa = get_irn_free_attr(a);
6060 const free_attr *pb = get_irn_free_attr(b);
6061 return (pa->where != pb->where) || (pa->type != pb->type);
6062 } /* node_cmp_attr_Free */
6064 /** Compares the attributes of two SymConst nodes. */
6065 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
6066 const symconst_attr *pa = get_irn_symconst_attr(a);
6067 const symconst_attr *pb = get_irn_symconst_attr(b);
6068 return (pa->kind != pb->kind)
6069 || (pa->sym.type_p != pb->sym.type_p)
6070 || (pa->tp != pb->tp);
6071 } /* node_cmp_attr_SymConst */
6073 /** Compares the attributes of two Call nodes. */
6074 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
6075 const call_attr *pa = get_irn_call_attr(a);
6076 const call_attr *pb = get_irn_call_attr(b);
6077 return (pa->type != pb->type)
6078 || (pa->tail_call != pb->tail_call);
6079 } /* node_cmp_attr_Call */
6081 /** Compares the attributes of two Sel nodes. */
6082 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
6083 const ir_entity *a_ent = get_Sel_entity(a);
6084 const ir_entity *b_ent = get_Sel_entity(b);
6087 (a_ent->kind != b_ent->kind) ||
6088 (a_ent->name != b_ent->name) ||
6089 (a_ent->owner != b_ent->owner) ||
6090 (a_ent->ld_name != b_ent->ld_name) ||
6091 (a_ent->type != b_ent->type);
6093 /* Matze: inlining of functions can produce 2 entities with same type,
6095 return a_ent != b_ent;
6096 } /* node_cmp_attr_Sel */
6098 /** Compares the attributes of two Phi nodes. */
6099 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
6100 /* we can only enter this function if both nodes have the same number of inputs,
6101 hence it is enough to check if one of them is a Phi0 */
6103 /* check the Phi0 pos attribute */
6104 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
6107 } /* node_cmp_attr_Phi */
6109 /** Compares the attributes of two Conv nodes. */
6110 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
6111 return get_Conv_strict(a) != get_Conv_strict(b);
6112 } /* node_cmp_attr_Conv */
6114 /** Compares the attributes of two Cast nodes. */
6115 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
6116 return get_Cast_type(a) != get_Cast_type(b);
6117 } /* node_cmp_attr_Cast */
6119 /** Compares the attributes of two Load nodes. */
6120 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
6121 if (get_Load_volatility(a) == volatility_is_volatile ||
6122 get_Load_volatility(b) == volatility_is_volatile)
6123 /* NEVER do CSE on volatile Loads */
6125 /* do not CSE Loads with different alignment. Be conservative. */
6126 if (get_Load_align(a) != get_Load_align(b))
6129 return get_Load_mode(a) != get_Load_mode(b);
6130 } /* node_cmp_attr_Load */
6132 /** Compares the attributes of two Store nodes. */
6133 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
6134 /* do not CSE Stores with different alignment. Be conservative. */
6135 if (get_Store_align(a) != get_Store_align(b))
6138 /* NEVER do CSE on volatile Stores */
6139 return (get_Store_volatility(a) == volatility_is_volatile ||
6140 get_Store_volatility(b) == volatility_is_volatile);
6141 } /* node_cmp_attr_Store */
6143 /** Compares two exception attributes */
6144 static int node_cmp_exception(ir_node *a, ir_node *b) {
6145 const except_attr *ea = get_irn_except_attr(a);
6146 const except_attr *eb = get_irn_except_attr(b);
6148 return ea->pin_state != eb->pin_state;
6151 #define node_cmp_attr_Bound node_cmp_exception
6153 /** Compares the attributes of two Div nodes. */
6154 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
6155 const divmod_attr *ma = get_irn_divmod_attr(a);
6156 const divmod_attr *mb = get_irn_divmod_attr(b);
6157 return ma->exc.pin_state != mb->exc.pin_state ||
6158 ma->resmode != mb->resmode ||
6159 ma->no_remainder != mb->no_remainder;
6160 } /* node_cmp_attr_Div */
6162 /** Compares the attributes of two DivMod nodes. */
6163 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
6164 const divmod_attr *ma = get_irn_divmod_attr(a);
6165 const divmod_attr *mb = get_irn_divmod_attr(b);
6166 return ma->exc.pin_state != mb->exc.pin_state ||
6167 ma->resmode != mb->resmode;
6168 } /* node_cmp_attr_DivMod */
6170 /** Compares the attributes of two Mod nodes. */
6171 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
6172 const divmod_attr *ma = get_irn_divmod_attr(a);
6173 const divmod_attr *mb = get_irn_divmod_attr(b);
6174 return ma->exc.pin_state != mb->exc.pin_state ||
6175 ma->resmode != mb->resmode;
6176 } /* node_cmp_attr_Mod */
6178 /** Compares the attributes of two Quot nodes. */
6179 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
6180 const divmod_attr *ma = get_irn_divmod_attr(a);
6181 const divmod_attr *mb = get_irn_divmod_attr(b);
6182 return ma->exc.pin_state != mb->exc.pin_state ||
6183 ma->resmode != mb->resmode;
6184 } /* node_cmp_attr_Quot */
6186 /** Compares the attributes of two Confirm nodes. */
6187 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
6188 /* no need to compare the bound, as this is a input */
6189 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
6190 } /* node_cmp_attr_Confirm */
6192 /** Compares the attributes of two Builtin nodes. */
6193 static int node_cmp_attr_Builtin(ir_node *a, ir_node *b) {
6194 const builtin_attr *ma = get_irn_builtin_attr(a);
6195 const builtin_attr *mb = get_irn_builtin_attr(b);
6197 /* no need to compare the type, equal kind means equal type */
6198 return ma->kind != mb->kind;
6199 } /* node_cmp_attr_Builtin */
6201 /** Compares the attributes of two ASM nodes. */
6202 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
6204 const ir_asm_constraint *ca;
6205 const ir_asm_constraint *cb;
6208 if (get_ASM_text(a) != get_ASM_text(b))
6211 /* Should we really check the constraints here? Should be better, but is strange. */
6212 n = get_ASM_n_input_constraints(a);
6213 if (n != get_ASM_n_input_constraints(b))
6216 ca = get_ASM_input_constraints(a);
6217 cb = get_ASM_input_constraints(b);
6218 for (i = 0; i < n; ++i) {
6219 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
6223 n = get_ASM_n_output_constraints(a);
6224 if (n != get_ASM_n_output_constraints(b))
6227 ca = get_ASM_output_constraints(a);
6228 cb = get_ASM_output_constraints(b);
6229 for (i = 0; i < n; ++i) {
6230 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
6234 n = get_ASM_n_clobbers(a);
6235 if (n != get_ASM_n_clobbers(b))
6238 cla = get_ASM_clobbers(a);
6239 clb = get_ASM_clobbers(b);
6240 for (i = 0; i < n; ++i) {
6241 if (cla[i] != clb[i])
6245 } /* node_cmp_attr_ASM */
6247 /** Compares the inexistent attributes of two Dummy nodes. */
6248 static int node_cmp_attr_Dummy(ir_node *a, ir_node *b)
6256 * Set the default node attribute compare operation for an ir_op_ops.
6258 * @param code the opcode for the default operation
6259 * @param ops the operations initialized
6264 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
6268 ops->node_cmp_attr = node_cmp_attr_##a; \
6301 } /* firm_set_default_node_cmp_attr */
6304 * Compare function for two nodes in the value table. Gets two
6305 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6307 int identities_cmp(const void *elt, const void *key) {
6308 ir_node *a = (ir_node *)elt;
6309 ir_node *b = (ir_node *)key;
6312 if (a == b) return 0;
6314 if ((get_irn_op(a) != get_irn_op(b)) ||
6315 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6317 /* compare if a's in and b's in are of equal length */
6318 irn_arity_a = get_irn_intra_arity(a);
6319 if (irn_arity_a != get_irn_intra_arity(b))
6322 if (get_irn_pinned(a) == op_pin_state_pinned) {
6323 /* for pinned nodes, the block inputs must be equal */
6324 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
6326 } else if (! get_opt_global_cse()) {
6327 /* for block-local CSE both nodes must be in the same MacroBlock */
6328 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
6332 /* compare a->in[0..ins] with b->in[0..ins] */
6333 for (i = 0; i < irn_arity_a; ++i) {
6334 ir_node *pred_a = get_irn_intra_n(a, i);
6335 ir_node *pred_b = get_irn_intra_n(b, i);
6336 if (pred_a != pred_b) {
6337 /* if both predecessors are CSE neutral they might be different */
6338 if (!is_irn_cse_neutral(pred_a) || !is_irn_cse_neutral(pred_b))
6344 * here, we already now that the nodes are identical except their
6347 if (a->op->ops.node_cmp_attr)
6348 return a->op->ops.node_cmp_attr(a, b);
6351 } /* identities_cmp */
6354 * Calculate a hash value of a node.
6356 * @param node The IR-node
6358 unsigned ir_node_hash(const ir_node *node) {
6359 return node->op->ops.hash(node);
6360 } /* ir_node_hash */
6363 pset *new_identities(void) {
6364 return new_pset(identities_cmp, N_IR_NODES);
6365 } /* new_identities */
6367 void del_identities(pset *value_table) {
6368 del_pset(value_table);
6369 } /* del_identities */
6371 /* Normalize a node by putting constants (and operands with larger
6372 * node index) on the right (operator side). */
6373 void ir_normalize_node(ir_node *n) {
6374 if (is_op_commutative(get_irn_op(n))) {
6375 ir_node *l = get_binop_left(n);
6376 ir_node *r = get_binop_right(n);
6378 /* For commutative operators perform a OP b == b OP a but keep
6379 * constants on the RIGHT side. This helps greatly in some
6380 * optimizations. Moreover we use the idx number to make the form
6382 if (!operands_are_normalized(l, r)) {
6383 set_binop_left(n, r);
6384 set_binop_right(n, l);
6388 } /* ir_normalize_node */
6391 * Update the nodes after a match in the value table. If both nodes have
6392 * the same MacroBlock but different Blocks, we must ensure that the node
6393 * with the dominating Block (the node that is near to the MacroBlock header
6394 * is stored in the table.
6395 * Because a MacroBlock has only one "non-exception" flow, we don't need
6396 * dominance info here: We known, that one block must dominate the other and
6397 * following the only block input will allow to find it.
6399 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
6400 ir_node *known_blk, *new_block, *block, *mbh;
6402 if (get_opt_global_cse()) {
6403 /* Block inputs are meaning less */
6406 known_blk = get_irn_n(known_irn, -1);
6407 new_block = get_irn_n(new_ir_node, -1);
6408 if (known_blk == new_block) {
6409 /* already in the same block */
6413 * We expect the typical case when we built the graph. In that case, the
6414 * known_irn is already the upper one, so checking this should be faster.
6417 mbh = get_Block_MacroBlock(new_block);
6419 if (block == known_blk) {
6420 /* ok, we have found it: known_block dominates new_block as expected */
6425 * We have reached the MacroBlock header NOT founding
6426 * the known_block. new_block must dominate known_block.
6429 set_irn_n(known_irn, -1, new_block);
6432 assert(get_Block_n_cfgpreds(block) == 1);
6433 block = get_Block_cfgpred_block(block, 0);
6435 } /* update_value_table */
6438 * Return the canonical node computing the same value as n.
6439 * Looks up the node in a hash table, enters it in the table
6440 * if it isn't there yet.
6442 * @param value_table the HashSet containing all nodes in the
6444 * @param n the node to look up
6446 * @return a node that computes the same value as n or n if no such
6447 * node could be found
6449 ir_node *identify_remember(pset *value_table, ir_node *n) {
6452 if (!value_table) return n;
6454 ir_normalize_node(n);
6455 /* lookup or insert in hash table with given hash key. */
6456 nn = pset_insert(value_table, n, ir_node_hash(n));
6459 update_known_irn(nn, n);
6461 /* n is reachable again */
6462 edges_node_revival(nn, get_irn_irg(nn));
6466 } /* identify_remember */
6469 * During construction we set the op_pin_state_pinned flag in the graph right when the
6470 * optimization is performed. The flag turning on procedure global cse could
6471 * be changed between two allocations. This way we are safe.
6473 * @param value_table The value table
6474 * @param n The node to lookup
6476 static inline ir_node *identify_cons(pset *value_table, ir_node *n) {
6479 n = identify_remember(value_table, n);
6480 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
6481 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6483 } /* identify_cons */
6485 /* Add a node to the identities value table. */
6486 void add_identities(pset *value_table, ir_node *node) {
6487 if (get_opt_cse() && is_no_Block(node))
6488 identify_remember(value_table, node);
6489 } /* add_identities */
6491 /* Visit each node in the value table of a graph. */
6492 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
6494 ir_graph *rem = current_ir_graph;
6496 current_ir_graph = irg;
6497 foreach_pset(irg->value_table, node)
6499 current_ir_graph = rem;
6500 } /* visit_all_identities */
6503 * Garbage in, garbage out. If a node has a dead input, i.e., the
6504 * Bad node is input to the node, return the Bad node.
6506 static ir_node *gigo(ir_node *node) {
6508 ir_op *op = get_irn_op(node);
6510 /* remove garbage blocks by looking at control flow that leaves the block
6511 and replacing the control flow by Bad. */
6512 if (get_irn_mode(node) == mode_X) {
6513 ir_node *block = get_nodes_block(skip_Proj(node));
6515 /* Don't optimize nodes in immature blocks. */
6516 if (!get_Block_matured(block))
6518 /* Don't optimize End, may have Bads. */
6519 if (op == op_End) return node;
6521 if (is_Block(block)) {
6522 if (is_Block_dead(block)) {
6523 /* control flow from dead block is dead */
6527 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
6528 if (!is_Bad(get_irn_n(block, i)))
6532 ir_graph *irg = get_irn_irg(block);
6533 /* the start block is never dead */
6534 if (block != get_irg_start_block(irg)
6535 && block != get_irg_end_block(irg)) {
6537 * Do NOT kill control flow without setting
6538 * the block to dead of bad things can happen:
6539 * We get a Block that is not reachable be irg_block_walk()
6540 * but can be found by irg_walk()!
6542 set_Block_dead(block);
6549 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6550 blocks predecessors is dead. */
6551 if (op != op_Block && op != op_Phi && op != op_Tuple) {
6552 irn_arity = get_irn_arity(node);
6555 * Beware: we can only read the block of a non-floating node.
6557 if (is_irn_pinned_in_irg(node) &&
6558 is_Block_dead(get_nodes_block(skip_Proj(node))))
6561 for (i = 0; i < irn_arity; i++) {
6562 ir_node *pred = get_irn_n(node, i);
6567 /* Propagating Unknowns here seems to be a bad idea, because
6568 sometimes we need a node as a input and did not want that
6570 However, it might be useful to move this into a later phase
6571 (if you think that optimizing such code is useful). */
6572 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
6573 return new_Unknown(get_irn_mode(node));
6578 /* With this code we violate the agreement that local_optimize
6579 only leaves Bads in Block, Phi and Tuple nodes. */
6580 /* If Block has only Bads as predecessors it's garbage. */
6581 /* If Phi has only Bads as predecessors it's garbage. */
6582 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
6583 irn_arity = get_irn_arity(node);
6584 for (i = 0; i < irn_arity; i++) {
6585 if (!is_Bad(get_irn_n(node, i))) break;
6587 if (i == irn_arity) node = new_Bad();
6594 * These optimizations deallocate nodes from the obstack.
6595 * It can only be called if it is guaranteed that no other nodes
6596 * reference this one, i.e., right after construction of a node.
6598 * @param n The node to optimize
6600 * current_ir_graph must be set to the graph of the node!
6602 ir_node *optimize_node(ir_node *n) {
6605 ir_opcode iro = get_irn_opcode(n);
6607 /* Always optimize Phi nodes: part of the construction. */
6608 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6610 /* constant expression evaluation / constant folding */
6611 if (get_opt_constant_folding()) {
6612 /* neither constants nor Tuple values can be evaluated */
6613 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6614 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6615 int old_fp_mode = tarval_fp_ops_enabled();
6617 tarval_enable_fp_ops(! (fp_model & fp_no_float_fold));
6619 /* try to evaluate */
6620 tv = computed_value(n);
6621 if (tv != tarval_bad) {
6623 ir_type *old_tp = get_irn_type(n);
6624 int i, arity = get_irn_arity(n);
6628 * Try to recover the type of the new expression.
6630 for (i = 0; i < arity && !old_tp; ++i)
6631 old_tp = get_irn_type(get_irn_n(n, i));
6634 * we MUST copy the node here temporary, because it's still needed
6635 * for DBG_OPT_CSTEVAL
6637 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6638 oldn = alloca(node_size);
6640 memcpy(oldn, n, node_size);
6641 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6643 /* ARG, copy the in array, we need it for statistics */
6644 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6646 /* note the inplace edges module */
6647 edges_node_deleted(n, current_ir_graph);
6649 /* evaluation was successful -- replace the node. */
6650 irg_kill_node(current_ir_graph, n);
6653 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6654 set_Const_type(nw, old_tp);
6655 DBG_OPT_CSTEVAL(oldn, nw);
6656 tarval_enable_fp_ops(old_fp_mode);
6659 tarval_enable_fp_ops(old_fp_mode);
6663 /* remove unnecessary nodes */
6664 if (get_opt_algebraic_simplification() ||
6665 (iro == iro_Phi) || /* always optimize these nodes. */
6667 (iro == iro_Proj) ||
6668 (iro == iro_Block) ) /* Flags tested local. */
6669 n = equivalent_node(n);
6671 /* Common Subexpression Elimination.
6673 * Checks whether n is already available.
6674 * The block input is used to distinguish different subexpressions. Right
6675 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6676 * subexpressions within a block.
6679 n = identify_cons(current_ir_graph->value_table, n);
6682 edges_node_deleted(oldn, current_ir_graph);
6684 /* We found an existing, better node, so we can deallocate the old node. */
6685 irg_kill_node(current_ir_graph, oldn);
6689 /* Some more constant expression evaluation that does not allow to
6691 iro = get_irn_opcode(n);
6692 if (get_opt_algebraic_simplification() ||
6693 (iro == iro_Cond) ||
6694 (iro == iro_Proj)) /* Flags tested local. */
6695 n = transform_node(n);
6697 /* Remove nodes with dead (Bad) input.
6698 Run always for transformation induced Bads. */
6701 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6702 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6704 n = identify_remember(current_ir_graph->value_table, o);
6710 } /* optimize_node */
6714 * These optimizations never deallocate nodes (in place). This can cause dead
6715 * nodes lying on the obstack. Remove these by a dead node elimination,
6716 * i.e., a copying garbage collection.
6718 ir_node *optimize_in_place_2(ir_node *n) {
6721 ir_opcode iro = get_irn_opcode(n);
6723 if (!get_opt_optimize() && !is_Phi(n)) return n;
6725 /* constant expression evaluation / constant folding */
6726 if (get_opt_constant_folding()) {
6727 /* neither constants nor Tuple values can be evaluated */
6728 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6729 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6730 int old_fp_mode = tarval_fp_ops_enabled();
6732 tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6733 /* try to evaluate */
6734 tv = computed_value(n);
6735 if (tv != tarval_bad) {
6736 /* evaluation was successful -- replace the node. */
6737 ir_type *old_tp = get_irn_type(n);
6738 int i, arity = get_irn_arity(n);
6741 * Try to recover the type of the new expression.
6743 for (i = 0; i < arity && !old_tp; ++i)
6744 old_tp = get_irn_type(get_irn_n(n, i));
6748 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6749 set_Const_type(n, old_tp);
6751 DBG_OPT_CSTEVAL(oldn, n);
6752 tarval_enable_fp_ops(old_fp_mode);
6755 tarval_enable_fp_ops(old_fp_mode);
6759 /* remove unnecessary nodes */
6760 if (get_opt_constant_folding() ||
6761 (iro == iro_Phi) || /* always optimize these nodes. */
6762 (iro == iro_Id) || /* ... */
6763 (iro == iro_Proj) || /* ... */
6764 (iro == iro_Block) ) /* Flags tested local. */
6765 n = equivalent_node(n);
6767 /** common subexpression elimination **/
6768 /* Checks whether n is already available. */
6769 /* The block input is used to distinguish different subexpressions. Right
6770 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6771 subexpressions within a block. */
6772 if (get_opt_cse()) {
6774 n = identify_remember(current_ir_graph->value_table, o);
6779 /* Some more constant expression evaluation. */
6780 iro = get_irn_opcode(n);
6781 if (get_opt_constant_folding() ||
6782 (iro == iro_Cond) ||
6783 (iro == iro_Proj)) /* Flags tested local. */
6784 n = transform_node(n);
6786 /* Remove nodes with dead (Bad) input.
6787 Run always for transformation induced Bads. */
6790 /* Now we can verify the node, as it has no dead inputs any more. */
6793 /* Now we have a legal, useful node. Enter it in hash table for cse.
6794 Blocks should be unique anyways. (Except the successor of start:
6795 is cse with the start block!) */
6796 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6798 n = identify_remember(current_ir_graph->value_table, o);
6804 } /* optimize_in_place_2 */
6807 * Wrapper for external use, set proper status bits after optimization.
6809 ir_node *optimize_in_place(ir_node *n) {
6810 /* Handle graph state */
6811 assert(get_irg_phase_state(current_ir_graph) != phase_building);
6813 if (get_opt_global_cse())
6814 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6815 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
6816 set_irg_outs_inconsistent(current_ir_graph);
6818 /* FIXME: Maybe we could also test whether optimizing the node can
6819 change the control graph. */
6820 set_irg_doms_inconsistent(current_ir_graph);
6821 return optimize_in_place_2(n);
6822 } /* optimize_in_place */
6825 * Calculate a hash value of a Const node.
6827 static unsigned hash_Const(const ir_node *node) {
6830 /* special value for const, as they only differ in their tarval. */
6831 h = HASH_PTR(node->attr.con.tv);
6837 * Calculate a hash value of a SymConst node.
6839 static unsigned hash_SymConst(const ir_node *node) {
6842 /* all others are pointers */
6843 h = HASH_PTR(node->attr.symc.sym.type_p);
6846 } /* hash_SymConst */
6849 * Set the default hash operation in an ir_op_ops.
6851 * @param code the opcode for the default operation
6852 * @param ops the operations initialized
6857 static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops)
6861 ops->hash = hash_##a; \
6864 /* hash function already set */
6865 if (ops->hash != NULL)
6872 /* use input/mode default hash if no function was given */
6873 ops->hash = firm_default_hash;
6881 * Sets the default operation for an ir_ops.
6883 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
6884 ops = firm_set_default_hash(code, ops);
6885 ops = firm_set_default_computed_value(code, ops);
6886 ops = firm_set_default_equivalent_node(code, ops);
6887 ops = firm_set_default_transform_node(code, ops);
6888 ops = firm_set_default_node_cmp_attr(code, ops);
6889 ops = firm_set_default_get_type(code, ops);
6890 ops = firm_set_default_get_type_attr(code, ops);
6891 ops = firm_set_default_get_entity_attr(code, ops);
6894 } /* firm_set_default_operations */