2 * Copyright (C) 1995-2011 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
32 #include "irgraph_t.h"
33 #include "iredges_t.h"
39 #include "iroptimize.h"
41 #include "dbginfo_t.h"
42 #include "iropt_dbg.h"
47 #include "opt_polymorphy.h"
52 #include "firm_types.h"
53 #include "bitfiddle.h"
56 /* Make types visible to allow most efficient access */
60 * Returns the tarval of a Const node or tarval_bad for all other nodes.
62 static ir_tarval *default_value_of(const ir_node *n)
65 return get_Const_tarval(n); /* might return tarval_bad */
70 value_of_func value_of_ptr = default_value_of;
72 /* * Set a new value_of function. */
73 void set_value_of_func(value_of_func func)
78 value_of_ptr = default_value_of;
82 * Return the value of a Constant.
84 static ir_tarval *computed_value_Const(const ir_node *n)
86 return get_Const_tarval(n);
87 } /* computed_value_Const */
90 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
92 static ir_tarval *computed_value_SymConst(const ir_node *n)
97 switch (get_SymConst_kind(n)) {
98 case symconst_type_size:
99 type = get_SymConst_type(n);
100 if (get_type_state(type) == layout_fixed)
101 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
103 case symconst_type_align:
104 type = get_SymConst_type(n);
105 if (get_type_state(type) == layout_fixed)
106 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
108 case symconst_ofs_ent:
109 ent = get_SymConst_entity(n);
110 type = get_entity_owner(ent);
111 if (get_type_state(type) == layout_fixed)
112 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
118 } /* computed_value_SymConst */
121 * Return the value of an Add.
123 static ir_tarval *computed_value_Add(const ir_node *n)
125 ir_node *a = get_Add_left(n);
126 ir_node *b = get_Add_right(n);
128 ir_tarval *ta = value_of(a);
129 ir_tarval *tb = value_of(b);
131 if ((ta != tarval_bad) && (tb != tarval_bad))
132 return tarval_add(ta, tb);
135 } /* computed_value_Add */
138 * Return the value of a Sub.
139 * Special case: a - a
141 static ir_tarval *computed_value_Sub(const ir_node *n)
143 ir_mode *mode = get_irn_mode(n);
144 ir_node *a = get_Sub_left(n);
145 ir_node *b = get_Sub_right(n);
150 if (! mode_is_float(mode)) {
153 return get_mode_null(mode);
159 if ((ta != tarval_bad) && (tb != tarval_bad))
160 return tarval_sub(ta, tb, mode);
163 } /* computed_value_Sub */
166 * Return the value of a Carry.
167 * Special : a op 0, 0 op b
169 static ir_tarval *computed_value_Carry(const ir_node *n)
171 ir_node *a = get_binop_left(n);
172 ir_node *b = get_binop_right(n);
173 ir_mode *m = get_irn_mode(n);
174 ir_tarval *ta = value_of(a);
175 ir_tarval *tb = value_of(b);
177 if ((ta != tarval_bad) && (tb != tarval_bad)) {
179 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
181 if (tarval_is_null(ta) || tarval_is_null(tb))
182 return get_mode_null(m);
185 } /* computed_value_Carry */
188 * Return the value of a Borrow.
191 static ir_tarval *computed_value_Borrow(const ir_node *n)
193 ir_node *a = get_binop_left(n);
194 ir_node *b = get_binop_right(n);
195 ir_mode *m = get_irn_mode(n);
196 ir_tarval *ta = value_of(a);
197 ir_tarval *tb = value_of(b);
199 if ((ta != tarval_bad) && (tb != tarval_bad)) {
200 return tarval_cmp(ta, tb) == ir_relation_less ? get_mode_one(m) : get_mode_null(m);
201 } else if (tarval_is_null(ta)) {
202 return get_mode_null(m);
205 } /* computed_value_Borrow */
208 * Return the value of an unary Minus.
210 static ir_tarval *computed_value_Minus(const ir_node *n)
212 ir_node *a = get_Minus_op(n);
213 ir_tarval *ta = value_of(a);
215 if (ta != tarval_bad)
216 return tarval_neg(ta);
219 } /* computed_value_Minus */
222 * Return the value of a Mul.
224 static ir_tarval *computed_value_Mul(const ir_node *n)
226 ir_node *a = get_Mul_left(n);
227 ir_node *b = get_Mul_right(n);
228 ir_tarval *ta = value_of(a);
229 ir_tarval *tb = value_of(b);
232 mode = get_irn_mode(n);
233 if (mode != get_irn_mode(a)) {
234 /* n * n = 2n bit multiplication */
235 ta = tarval_convert_to(ta, mode);
236 tb = tarval_convert_to(tb, mode);
239 if (ta != tarval_bad && tb != tarval_bad) {
240 return tarval_mul(ta, tb);
242 /* a * 0 != 0 if a == NaN or a == Inf */
243 if (!mode_is_float(mode)) {
244 /* a*0 = 0 or 0*b = 0 */
245 if (ta == get_mode_null(mode))
247 if (tb == get_mode_null(mode))
252 } /* computed_value_Mul */
255 * Return the value of an And.
256 * Special case: a & 0, 0 & b
258 static ir_tarval *computed_value_And(const ir_node *n)
260 ir_node *a = get_And_left(n);
261 ir_node *b = get_And_right(n);
262 ir_tarval *ta = value_of(a);
263 ir_tarval *tb = value_of(b);
265 if ((ta != tarval_bad) && (tb != tarval_bad)) {
266 return tarval_and (ta, tb);
268 if (tarval_is_null(ta)) return ta;
269 if (tarval_is_null(tb)) return tb;
272 } /* computed_value_And */
275 * Return the value of an Or.
276 * Special case: a | 1...1, 1...1 | b
278 static ir_tarval *computed_value_Or(const ir_node *n)
280 ir_node *a = get_Or_left(n);
281 ir_node *b = get_Or_right(n);
282 ir_tarval *ta = value_of(a);
283 ir_tarval *tb = value_of(b);
285 if ((ta != tarval_bad) && (tb != tarval_bad)) {
286 return tarval_or (ta, tb);
288 if (tarval_is_all_one(ta)) return ta;
289 if (tarval_is_all_one(tb)) return tb;
292 } /* computed_value_Or */
295 * Return the value of an Eor.
297 static ir_tarval *computed_value_Eor(const ir_node *n)
299 ir_node *a = get_Eor_left(n);
300 ir_node *b = get_Eor_right(n);
305 return get_mode_null(get_irn_mode(n));
310 if ((ta != tarval_bad) && (tb != tarval_bad)) {
311 return tarval_eor(ta, tb);
314 } /* computed_value_Eor */
317 * Return the value of a Not.
319 static ir_tarval *computed_value_Not(const ir_node *n)
321 ir_node *a = get_Not_op(n);
322 ir_tarval *ta = value_of(a);
324 if (ta != tarval_bad)
325 return tarval_not(ta);
328 } /* computed_value_Not */
331 * Return the value of a Shl.
333 static ir_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 ir_tarval *ta = value_of(a);
339 ir_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 ir_tarval *computed_value_Shr(const ir_node *n)
352 ir_node *a = get_Shr_left(n);
353 ir_node *b = get_Shr_right(n);
355 ir_tarval *ta = value_of(a);
356 ir_tarval *tb = value_of(b);
358 if ((ta != tarval_bad) && (tb != tarval_bad)) {
359 return tarval_shr(ta, tb);
362 } /* computed_value_Shr */
365 * Return the value of a Shrs.
367 static ir_tarval *computed_value_Shrs(const ir_node *n)
369 ir_node *a = get_Shrs_left(n);
370 ir_node *b = get_Shrs_right(n);
372 ir_tarval *ta = value_of(a);
373 ir_tarval *tb = value_of(b);
375 if ((ta != tarval_bad) && (tb != tarval_bad)) {
376 return tarval_shrs(ta, tb);
379 } /* computed_value_Shrs */
382 * Return the value of a Rotl.
384 static ir_tarval *computed_value_Rotl(const ir_node *n)
386 ir_node *a = get_Rotl_left(n);
387 ir_node *b = get_Rotl_right(n);
389 ir_tarval *ta = value_of(a);
390 ir_tarval *tb = value_of(b);
392 if ((ta != tarval_bad) && (tb != tarval_bad)) {
393 return tarval_rotl(ta, tb);
396 } /* computed_value_Rotl */
399 * Return the value of a Conv.
401 static ir_tarval *computed_value_Conv(const ir_node *n)
403 ir_node *a = get_Conv_op(n);
404 ir_tarval *ta = value_of(a);
406 if (ta != tarval_bad)
407 return tarval_convert_to(ta, get_irn_mode(n));
410 } /* computed_value_Conv */
413 * Calculate the value of a Mux: can be evaluated, if the
414 * sel and the right input are known.
416 static ir_tarval *computed_value_Mux(const ir_node *n)
418 ir_node *sel = get_Mux_sel(n);
419 ir_tarval *ts = value_of(sel);
421 if (ts == get_tarval_b_true()) {
422 ir_node *v = get_Mux_true(n);
425 else if (ts == get_tarval_b_false()) {
426 ir_node *v = get_Mux_false(n);
430 } /* computed_value_Mux */
433 * Calculate the value of a Confirm: can be evaluated,
434 * if it has the form Confirm(x, '=', Const).
436 static ir_tarval *computed_value_Confirm(const ir_node *n)
438 if (get_Confirm_relation(n) == ir_relation_equal) {
439 ir_tarval *tv = value_of(get_Confirm_bound(n));
440 if (tv != tarval_bad)
443 return value_of(get_Confirm_value(n));
444 } /* computed_value_Confirm */
447 * gives a (conservative) estimation of possible relation when comparing
450 ir_relation ir_get_possible_cmp_relations(const ir_node *left,
451 const ir_node *right)
453 ir_relation possible = ir_relation_true;
454 ir_tarval *tv_l = value_of(left);
455 ir_tarval *tv_r = value_of(right);
456 ir_mode *mode = get_irn_mode(left);
457 ir_tarval *min = mode == mode_b ? tarval_b_false : get_mode_min(mode);
458 ir_tarval *max = mode == mode_b ? tarval_b_true : get_mode_max(mode);
460 /* both values known - evaluate them */
461 if ((tv_l != tarval_bad) && (tv_r != tarval_bad)) {
462 possible = tarval_cmp(tv_l, tv_r);
463 /* we can return now, won't get any better */
466 /* a == a is never less or greater (but might be equal or unordered) */
468 possible &= ~ir_relation_less_greater;
469 /* unordered results only happen for float compares */
470 if (!mode_is_float(mode))
471 possible &= ~ir_relation_unordered;
472 /* values can never be less than the least representable number or
473 * greater than the greatest representable number */
475 possible &= ~ir_relation_greater;
477 possible &= ~ir_relation_less;
479 possible &= ~ir_relation_greater;
481 possible &= ~ir_relation_less;
482 /* maybe vrp can tell us more */
483 possible &= vrp_cmp(left, right);
484 /* Alloc nodes never return null (but throw an exception) */
485 if (is_Alloc(left) && tarval_is_null(tv_r))
486 possible &= ~ir_relation_equal;
492 * Return the value of a Cmp.
494 * The basic idea here is to determine which relations are possible and which
495 * one are definitely impossible.
497 static ir_tarval *computed_value_Cmp(const ir_node *cmp)
499 ir_node *left = get_Cmp_left(cmp);
500 ir_node *right = get_Cmp_right(cmp);
501 ir_relation possible = ir_get_possible_cmp_relations(left, right);
502 ir_relation relation = get_Cmp_relation(cmp);
504 /* if none of the requested relations is possible, return false */
505 if ((possible & relation) == ir_relation_false)
506 return tarval_b_false;
507 /* if possible relations are a subset of the requested ones return true */
508 if ((possible & ~relation) == ir_relation_false)
509 return tarval_b_true;
511 return computed_value_Cmp_Confirm(cmp, left, right, relation);
515 * Calculate the value of an integer Div.
516 * Special case: 0 / b
518 static ir_tarval *do_computed_value_Div(const ir_node *div)
520 const ir_node *a = get_Div_left(div);
521 const ir_node *b = get_Div_right(div);
522 const ir_mode *mode = get_Div_resmode(div);
523 ir_tarval *ta = value_of(a);
525 const ir_node *dummy;
527 /* cannot optimize 0 / b = 0 because of NaN */
528 if (!mode_is_float(mode)) {
529 if (tarval_is_null(ta) && value_not_zero(b, &dummy))
530 return ta; /* 0 / b == 0 if b != 0 */
533 if (ta != tarval_bad && tb != tarval_bad)
534 return tarval_div(ta, tb);
536 } /* do_computed_value_Div */
539 * Calculate the value of an integer Mod of two nodes.
540 * Special case: a % 1
542 static ir_tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b)
544 ir_tarval *ta = value_of(a);
545 ir_tarval *tb = value_of(b);
547 /* Compute a % 1 or c1 % c2 */
548 if (tarval_is_one(tb))
549 return get_mode_null(get_irn_mode(a));
550 if (ta != tarval_bad && tb != tarval_bad)
551 return tarval_mod(ta, tb);
553 } /* do_computed_value_Mod */
556 * Return the value of a Proj(Div).
558 static ir_tarval *computed_value_Proj_Div(const ir_node *n)
560 long proj_nr = get_Proj_proj(n);
561 if (proj_nr != pn_Div_res)
564 return do_computed_value_Div(get_Proj_pred(n));
565 } /* computed_value_Proj_Div */
568 * Return the value of a Proj(Mod).
570 static ir_tarval *computed_value_Proj_Mod(const ir_node *n)
572 long proj_nr = get_Proj_proj(n);
574 if (proj_nr == pn_Mod_res) {
575 const ir_node *mod = get_Proj_pred(n);
576 return do_computed_value_Mod(get_Mod_left(mod), get_Mod_right(mod));
579 } /* computed_value_Proj_Mod */
582 * Return the value of a Proj.
584 static ir_tarval *computed_value_Proj(const ir_node *proj)
586 ir_node *n = get_Proj_pred(proj);
588 if (n->op->ops.computed_value_Proj != NULL)
589 return n->op->ops.computed_value_Proj(proj);
591 } /* computed_value_Proj */
594 * If the parameter n can be computed, return its value, else tarval_bad.
595 * Performs constant folding.
597 * @param n The node this should be evaluated
599 ir_tarval *computed_value(const ir_node *n)
601 vrp_attr *vrp = vrp_get_info(n);
602 if (vrp && vrp->valid && tarval_cmp(vrp->bits_set, vrp->bits_not_set) == ir_relation_equal) {
603 return vrp->bits_set;
605 if (n->op->ops.computed_value)
606 return n->op->ops.computed_value(n);
608 } /* computed_value */
611 * Set the default computed_value evaluator in an ir_op_ops.
613 * @param code the opcode for the default operation
614 * @param ops the operations initialized
619 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
623 ops->computed_value = computed_value_##a; \
625 #define CASE_PROJ(a) \
627 ops->computed_value_Proj = computed_value_Proj_##a; \
662 } /* firm_set_default_computed_value */
665 * Returns a equivalent block for another block.
666 * If the block has only one predecessor, this is
667 * the equivalent one. If the only predecessor of a block is
668 * the block itself, this is a dead block.
670 * If both predecessors of a block are the branches of a binary
671 * Cond, the equivalent block is Cond's block.
673 * If all predecessors of a block are bad or lies in a dead
674 * block, the current block is dead as well.
676 * Note, that blocks are NEVER turned into Bad's, instead
677 * the dead_block flag is set. So, never test for is_Bad(block),
678 * always use is_dead_Block(block).
680 static ir_node *equivalent_node_Block(ir_node *n)
686 /* don't optimize dead or labeled blocks */
687 if (is_Block_dead(n) || has_Block_entity(n))
690 n_preds = get_Block_n_cfgpreds(n);
692 /* The Block constructor does not call optimize, but mature_immBlock()
693 calls the optimization. */
694 assert(get_Block_matured(n));
696 irg = get_irn_irg(n);
698 /* Straightening: a single entry Block following a single exit Block
699 can be merged, if it is not the Start block. */
700 /* !!! Beware, all Phi-nodes of n must have been optimized away.
701 This should be true, as the block is matured before optimize is called.
702 But what about Phi-cycles with the Phi0/Id that could not be resolved?
703 Remaining Phi nodes are just Ids. */
705 ir_node *pred = skip_Proj(get_Block_cfgpred(n, 0));
708 ir_node *predblock = get_nodes_block(pred);
709 if (predblock == oldn) {
710 /* Jmp jumps into the block it is in -- deal self cycle. */
711 n = set_Block_dead(n);
712 DBG_OPT_DEAD_BLOCK(oldn, n);
715 DBG_OPT_STG(oldn, n);
717 } else if (is_Cond(pred)) {
718 ir_node *predblock = get_nodes_block(pred);
719 if (predblock == oldn) {
720 /* Jmp jumps into the block it is in -- deal self cycle. */
721 n = set_Block_dead(n);
722 DBG_OPT_DEAD_BLOCK(oldn, n);
725 } else if (n_preds == 2) {
726 /* Test whether Cond jumps twice to this block
727 * The more general case which more than 2 predecessors is handles
728 * in optimize_cf(), we handle only this special case for speed here.
730 ir_node *a = get_Block_cfgpred(n, 0);
731 ir_node *b = get_Block_cfgpred(n, 1);
733 if (is_Proj(a) && is_Proj(b)) {
734 ir_node *cond = get_Proj_pred(a);
736 if (cond == get_Proj_pred(b) && is_Cond(cond) &&
737 get_irn_mode(get_Cond_selector(cond)) == mode_b) {
738 /* Also a single entry Block following a single exit Block. Phis have
739 twice the same operand and will be optimized away. */
740 n = get_nodes_block(cond);
741 DBG_OPT_IFSIM1(oldn, a, b, n);
744 } else if (get_opt_unreachable_code() &&
745 (n != get_irg_start_block(irg)) &&
746 (n != get_irg_end_block(irg))) {
749 /* If all inputs are dead, this block is dead too, except if it is
750 the start or end block. This is one step of unreachable code
752 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
753 ir_node *pred = get_Block_cfgpred(n, i);
756 if (is_Bad(pred)) continue;
757 pred_blk = get_nodes_block(skip_Proj(pred));
759 if (is_Block_dead(pred_blk)) continue;
762 /* really found a living input */
767 n = set_Block_dead(n);
768 DBG_OPT_DEAD_BLOCK(oldn, n);
773 } /* equivalent_node_Block */
776 * Returns a equivalent node for a Jmp, a Bad :-)
777 * Of course this only happens if the Block of the Jmp is dead.
779 static ir_node *equivalent_node_Jmp(ir_node *n)
783 /* unreachable code elimination */
784 if (is_Block_dead(get_nodes_block(n))) {
785 ir_graph *irg = get_irn_irg(n);
786 n = get_irg_bad(irg);
787 DBG_OPT_DEAD_BLOCK(oldn, n);
790 } /* equivalent_node_Jmp */
792 /** Raise is handled in the same way as Jmp. */
793 #define equivalent_node_Raise equivalent_node_Jmp
796 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
797 See transform_node_Proj_Cond(). */
800 * Optimize operations that are commutative and have neutral 0,
801 * so a op 0 = 0 op a = a.
803 static ir_node *equivalent_node_neutral_zero(ir_node *n)
807 ir_node *a = get_binop_left(n);
808 ir_node *b = get_binop_right(n);
813 /* After running compute_node there is only one constant predecessor.
814 Find this predecessors value and remember the other node: */
815 if ((tv = value_of(a)) != tarval_bad) {
817 } else if ((tv = value_of(b)) != tarval_bad) {
822 /* If this predecessors constant value is zero, the operation is
823 * unnecessary. Remove it.
825 * Beware: If n is a Add, the mode of on and n might be different
826 * which happens in this rare construction: NULL + 3.
827 * Then, a Conv would be needed which we cannot include here.
829 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
832 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
836 } /* equivalent_node_neutral_zero */
839 * Eor is commutative and has neutral 0.
841 static ir_node *equivalent_node_Eor(ir_node *n)
847 n = equivalent_node_neutral_zero(n);
848 if (n != oldn) return n;
851 b = get_Eor_right(n);
854 ir_node *aa = get_Eor_left(a);
855 ir_node *ab = get_Eor_right(a);
858 /* (a ^ b) ^ a -> b */
860 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
862 } else if (ab == b) {
863 /* (a ^ b) ^ b -> a */
865 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
870 ir_node *ba = get_Eor_left(b);
871 ir_node *bb = get_Eor_right(b);
874 /* a ^ (a ^ b) -> b */
876 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
878 } else if (bb == a) {
879 /* a ^ (b ^ a) -> b */
881 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
889 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
891 * The second one looks strange, but this construct
892 * is used heavily in the LCC sources :-).
894 * Beware: The Mode of an Add may be different than the mode of its
895 * predecessors, so we could not return a predecessors in all cases.
897 static ir_node *equivalent_node_Add(ir_node *n)
900 ir_node *left, *right;
901 ir_mode *mode = get_irn_mode(n);
903 n = equivalent_node_neutral_zero(n);
907 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
908 if (mode_is_float(mode)) {
909 ir_graph *irg = get_irn_irg(n);
910 if (get_irg_fp_model(irg) & fp_strict_algebraic)
914 left = get_Add_left(n);
915 right = get_Add_right(n);
918 if (get_Sub_right(left) == right) {
921 n = get_Sub_left(left);
922 if (mode == get_irn_mode(n)) {
923 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
929 if (get_Sub_right(right) == left) {
932 n = get_Sub_left(right);
933 if (mode == get_irn_mode(n)) {
934 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
940 } /* equivalent_node_Add */
943 * optimize operations that are not commutative but have neutral 0 on left,
946 static ir_node *equivalent_node_left_zero(ir_node *n)
950 ir_node *a = get_binop_left(n);
951 ir_node *b = get_binop_right(n);
952 ir_tarval *tb = value_of(b);
954 if (tarval_is_null(tb)) {
957 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
960 } /* equivalent_node_left_zero */
962 #define equivalent_node_Shl equivalent_node_left_zero
963 #define equivalent_node_Shr equivalent_node_left_zero
964 #define equivalent_node_Shrs equivalent_node_left_zero
965 #define equivalent_node_Rotl equivalent_node_left_zero
968 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
970 * The second one looks strange, but this construct
971 * is used heavily in the LCC sources :-).
973 * Beware: The Mode of a Sub may be different than the mode of its
974 * predecessors, so we could not return a predecessors in all cases.
976 static ir_node *equivalent_node_Sub(ir_node *n)
980 ir_mode *mode = get_irn_mode(n);
983 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
984 if (mode_is_float(mode)) {
985 ir_graph *irg = get_irn_irg(n);
986 if (get_irg_fp_model(irg) & fp_strict_algebraic)
990 b = get_Sub_right(n);
993 /* Beware: modes might be different */
994 if (tarval_is_null(tb)) {
995 ir_node *a = get_Sub_left(n);
996 if (mode == get_irn_mode(a)) {
999 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1003 } /* equivalent_node_Sub */
1007 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
1010 * -(-a) == a, but might overflow two times.
1011 * We handle it anyway here but the better way would be a
1012 * flag. This would be needed for Pascal for instance.
1014 static ir_node *equivalent_node_idempotent_unop(ir_node *n)
1017 ir_node *pred = get_unop_op(n);
1019 /* optimize symmetric unop */
1020 if (get_irn_op(pred) == get_irn_op(n)) {
1021 n = get_unop_op(pred);
1022 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
1025 } /* equivalent_node_idempotent_unop */
1027 /** Optimize Not(Not(x)) == x. */
1028 #define equivalent_node_Not equivalent_node_idempotent_unop
1030 /** -(-x) == x ??? Is this possible or can --x raise an
1031 out of bounds exception if min =! max? */
1032 #define equivalent_node_Minus equivalent_node_idempotent_unop
1035 * Optimize a * 1 = 1 * a = a.
1037 static ir_node *equivalent_node_Mul(ir_node *n)
1040 ir_node *a = get_Mul_left(n);
1042 /* we can handle here only the n * n = n bit cases */
1043 if (get_irn_mode(n) == get_irn_mode(a)) {
1044 ir_node *b = get_Mul_right(n);
1048 * Mul is commutative and has again an other neutral element.
1049 * Constants are place right, so check this case first.
1052 if (tarval_is_one(tv)) {
1054 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1057 if (tarval_is_one(tv)) {
1059 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1064 } /* equivalent_node_Mul */
1067 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1069 static ir_node *equivalent_node_Or(ir_node *n)
1073 ir_node *a = get_Or_left(n);
1074 ir_node *b = get_Or_right(n);
1078 n = a; /* idempotence */
1079 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1082 /* constants are normalized to right, check this side first */
1084 if (tarval_is_null(tv)) {
1086 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1090 if (tarval_is_null(tv)) {
1092 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1097 } /* equivalent_node_Or */
1100 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1102 static ir_node *equivalent_node_And(ir_node *n)
1106 ir_node *a = get_And_left(n);
1107 ir_node *b = get_And_right(n);
1111 n = a; /* idempotence */
1112 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1115 /* constants are normalized to right, check this side first */
1117 if (tarval_is_all_one(tv)) {
1119 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1122 if (tv != get_tarval_bad()) {
1123 ir_mode *mode = get_irn_mode(n);
1124 if (!mode_is_signed(mode) && is_Conv(a)) {
1125 ir_node *convop = get_Conv_op(a);
1126 ir_mode *convopmode = get_irn_mode(convop);
1127 if (!mode_is_signed(convopmode)) {
1128 if (tarval_is_all_one(tarval_convert_to(tv, convopmode))) {
1129 /* Conv(X) & all_one(mode(X)) = Conv(X) */
1131 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1138 if (tarval_is_all_one(tv)) {
1140 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1144 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1147 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1152 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1155 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1160 } /* equivalent_node_And */
1163 * Try to remove useless Conv's:
1165 static ir_node *equivalent_node_Conv(ir_node *n)
1168 ir_node *a = get_Conv_op(n);
1170 ir_mode *n_mode = get_irn_mode(n);
1171 ir_mode *a_mode = get_irn_mode(a);
1174 if (n_mode == a_mode) { /* No Conv necessary */
1175 if (get_Conv_strict(n)) {
1178 /* neither Minus nor Confirm change the precision,
1179 so we can "look-through" */
1182 p = get_Minus_op(p);
1183 } else if (is_Confirm(p)) {
1184 p = get_Confirm_value(p);
1190 if (is_Conv(p) && get_Conv_strict(p)) {
1191 /* we known already, that a_mode == n_mode, and neither
1192 Minus change the mode, so the second Conv
1194 assert(get_irn_mode(p) == n_mode);
1196 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1200 ir_node *pred = get_Proj_pred(p);
1201 if (is_Load(pred)) {
1202 /* Loads always return with the exact precision of n_mode */
1203 assert(get_Load_mode(pred) == n_mode);
1205 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1208 if (is_Proj(pred) && get_Proj_proj(pred) == pn_Start_T_args) {
1209 pred = get_Proj_pred(pred);
1210 if (is_Start(pred)) {
1211 /* Arguments always return with the exact precision,
1212 as strictConv's are place before Call -- if the
1213 caller was compiled with the same setting.
1214 Otherwise, the semantics is probably still right. */
1215 assert(get_irn_mode(p) == n_mode);
1217 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1223 /* special case: the immediate predecessor is also a Conv */
1224 if (! get_Conv_strict(a)) {
1225 /* first one is not strict, kick it */
1227 a_mode = get_irn_mode(a);
1231 /* else both are strict conv, second is superfluous */
1233 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1238 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1241 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1242 ir_node *b = get_Conv_op(a);
1243 ir_mode *b_mode = get_irn_mode(b);
1245 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1246 /* both are strict conv */
1247 if (smaller_mode(a_mode, n_mode)) {
1248 /* both are strict, but the first is smaller, so
1249 the second cannot remove more precision, remove the
1251 set_Conv_strict(n, 0);
1254 if (n_mode == b_mode) {
1255 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1256 if (n_mode == mode_b) {
1257 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1258 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1260 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1261 if (values_in_mode(b_mode, a_mode)) {
1262 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1263 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1268 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1269 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1270 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1271 unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode);
1273 if (float_mantissa >= int_mantissa) {
1275 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1280 if (smaller_mode(b_mode, a_mode)) {
1281 if (get_Conv_strict(n))
1282 set_Conv_strict(b, 1);
1283 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1284 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1291 } /* equivalent_node_Conv */
1294 * - fold Phi-nodes, iff they have only one predecessor except
1297 static ir_node *equivalent_node_Phi(ir_node *n)
1303 ir_node *first_val = NULL; /* to shutup gcc */
1305 if (!get_opt_optimize() &&
1306 get_irg_phase_state(get_irn_irg(n)) != phase_building)
1309 n_preds = get_Phi_n_preds(n);
1311 block = get_nodes_block(n);
1313 if (is_Block_dead(block)) {
1314 ir_graph *irg = get_irn_irg(n);
1315 return get_irg_bad(irg);
1318 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1320 /* Find first non-self-referencing input */
1321 for (i = 0; i < n_preds; ++i) {
1322 first_val = get_Phi_pred(n, i);
1323 if ( (first_val != n) /* not self pointer */
1325 /* BEWARE: when the if is changed to 1, Phis will ignore their Bad
1326 * predecessors. Then, Phi nodes in unreachable code might be removed,
1327 * causing nodes pointing to themselev (Adds for instance).
1328 * This is really bad and causes endless recursion on several
1329 * code pathes, so we do NOT optimize such code.
1330 * This is not that bad as it sounds, optimize_cf() removes bad control flow
1331 * (and bad Phi predecessors), so live code is optimized later.
1333 && (! is_Bad(get_Block_cfgpred(block, i)))
1335 ) { /* value not dead */
1336 break; /* then found first value. */
1341 ir_graph *irg = get_irn_irg(n);
1342 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1343 return get_irg_bad(irg);
1346 /* search for rest of inputs, determine if any of these
1347 are non-self-referencing */
1348 while (++i < n_preds) {
1349 ir_node *scnd_val = get_Phi_pred(n, i);
1350 if ( (scnd_val != n)
1351 && (scnd_val != first_val)
1354 && (! is_Bad(get_Block_cfgpred(block, i)))
1361 if (i >= n_preds && !is_Dummy(first_val)) {
1362 /* Fold, if no multiple distinct non-self-referencing inputs */
1364 DBG_OPT_PHI(oldn, n);
1367 } /* equivalent_node_Phi */
1370 * Several optimizations:
1371 * - fold Sync-nodes, iff they have only one predecessor except
1374 static ir_node *equivalent_node_Sync(ir_node *n)
1376 int arity = get_Sync_n_preds(n);
1379 for (i = 0; i < arity;) {
1380 ir_node *pred = get_Sync_pred(n, i);
1383 /* Remove Bad predecessors */
1390 /* Remove duplicate predecessors */
1396 if (get_Sync_pred(n, j) == pred) {
1405 ir_graph *irg = get_irn_irg(n);
1406 return get_irg_bad(irg);
1408 if (arity == 1) return get_Sync_pred(n, 0);
1410 } /* equivalent_node_Sync */
1413 * Optimize Proj(Tuple).
1415 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj)
1417 ir_node *oldn = proj;
1418 ir_node *tuple = get_Proj_pred(proj);
1420 /* Remove the Tuple/Proj combination. */
1421 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1422 DBG_OPT_TUPLE(oldn, tuple, proj);
1425 } /* equivalent_node_Proj_Tuple */
1428 * Optimize a / 1 = a.
1430 static ir_node *equivalent_node_Proj_Div(ir_node *proj)
1432 ir_node *oldn = proj;
1433 ir_node *div = get_Proj_pred(proj);
1434 ir_node *b = get_Div_right(div);
1435 ir_tarval *tb = value_of(b);
1437 /* Div is not commutative. */
1438 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1439 switch (get_Proj_proj(proj)) {
1441 proj = get_Div_mem(div);
1442 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1446 proj = get_Div_left(div);
1447 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1451 /* we cannot replace the exception Proj's here, this is done in
1452 transform_node_Proj_Div() */
1457 } /* equivalent_node_Proj_Div */
1460 * Optimize CopyB(mem, x, x) into a Nop.
1462 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj)
1464 ir_node *oldn = proj;
1465 ir_node *copyb = get_Proj_pred(proj);
1466 ir_node *a = get_CopyB_dst(copyb);
1467 ir_node *b = get_CopyB_src(copyb);
1470 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1471 switch (get_Proj_proj(proj)) {
1473 proj = get_CopyB_mem(copyb);
1474 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1477 case pn_CopyB_X_except: {
1478 ir_graph *irg = get_irn_irg(proj);
1479 DBG_OPT_EXC_REM(proj);
1480 proj = get_irg_bad(irg);
1486 } /* equivalent_node_Proj_CopyB */
1489 * Optimize Bounds(idx, idx, upper) into idx.
1491 static ir_node *equivalent_node_Proj_Bound(ir_node *proj)
1493 ir_node *oldn = proj;
1494 ir_node *bound = get_Proj_pred(proj);
1495 ir_node *idx = get_Bound_index(bound);
1496 ir_node *pred = skip_Proj(idx);
1499 if (idx == get_Bound_lower(bound))
1501 else if (is_Bound(pred)) {
1503 * idx was Bounds checked previously, it is still valid if
1504 * lower <= pred_lower && pred_upper <= upper.
1506 ir_node *lower = get_Bound_lower(bound);
1507 ir_node *upper = get_Bound_upper(bound);
1508 if (get_Bound_lower(pred) == lower &&
1509 get_Bound_upper(pred) == upper) {
1511 * One could expect that we simply return the previous
1512 * Bound here. However, this would be wrong, as we could
1513 * add an exception Proj to a new location then.
1514 * So, we must turn in into a tuple.
1520 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1521 switch (get_Proj_proj(proj)) {
1523 DBG_OPT_EXC_REM(proj);
1524 proj = get_Bound_mem(bound);
1526 case pn_Bound_X_except: {
1527 ir_graph *irg = get_irn_irg(proj);
1528 DBG_OPT_EXC_REM(proj);
1529 proj = get_irg_bad(irg);
1534 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1537 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1542 } /* equivalent_node_Proj_Bound */
1545 * Optimize an Exception Proj(Load) with a non-null address.
1547 static ir_node *equivalent_node_Proj_Load(ir_node *proj)
1549 if (get_opt_ldst_only_null_ptr_exceptions()) {
1550 if (get_irn_mode(proj) == mode_X) {
1551 ir_node *load = get_Proj_pred(proj);
1553 /* get the Load address */
1554 const ir_node *addr = get_Load_ptr(load);
1555 const ir_node *confirm;
1557 if (value_not_null(addr, &confirm)) {
1558 if (get_Proj_proj(proj) == pn_Load_X_except) {
1559 ir_graph *irg = get_irn_irg(proj);
1560 DBG_OPT_EXC_REM(proj);
1561 return get_irg_bad(irg);
1567 } /* equivalent_node_Proj_Load */
1570 * Optimize an Exception Proj(Store) with a non-null address.
1572 static ir_node *equivalent_node_Proj_Store(ir_node *proj)
1574 if (get_opt_ldst_only_null_ptr_exceptions()) {
1575 if (get_irn_mode(proj) == mode_X) {
1576 ir_node *store = get_Proj_pred(proj);
1578 /* get the load/store address */
1579 const ir_node *addr = get_Store_ptr(store);
1580 const ir_node *confirm;
1582 if (value_not_null(addr, &confirm)) {
1583 if (get_Proj_proj(proj) == pn_Store_X_except) {
1584 ir_graph *irg = get_irn_irg(proj);
1585 DBG_OPT_EXC_REM(proj);
1586 return get_irg_bad(irg);
1592 } /* equivalent_node_Proj_Store */
1595 * Does all optimizations on nodes that must be done on its Projs
1596 * because of creating new nodes.
1598 static ir_node *equivalent_node_Proj(ir_node *proj)
1600 ir_node *n = get_Proj_pred(proj);
1602 if (get_irn_mode(proj) == mode_X) {
1603 if (is_Block_dead(get_nodes_block(n))) {
1604 /* Remove dead control flow -- early gigo(). */
1605 ir_graph *irg = get_irn_irg(proj);
1606 return get_irg_bad(irg);
1609 if (n->op->ops.equivalent_node_Proj)
1610 return n->op->ops.equivalent_node_Proj(proj);
1612 } /* equivalent_node_Proj */
1617 static ir_node *equivalent_node_Id(ir_node *n)
1625 DBG_OPT_ID(oldn, n);
1627 } /* equivalent_node_Id */
1632 static ir_node *equivalent_node_Mux(ir_node *n)
1634 ir_node *oldn = n, *sel = get_Mux_sel(n);
1636 ir_tarval *ts = value_of(sel);
1638 /* Mux(true, f, t) == t */
1639 if (ts == tarval_b_true) {
1640 n = get_Mux_true(n);
1641 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1644 /* Mux(false, f, t) == f */
1645 if (ts == tarval_b_false) {
1646 n = get_Mux_false(n);
1647 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1650 n_t = get_Mux_true(n);
1651 n_f = get_Mux_false(n);
1653 /* Mux(v, x, T) == x */
1654 if (is_Unknown(n_f)) {
1656 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1659 /* Mux(v, T, x) == x */
1660 if (is_Unknown(n_t)) {
1662 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1666 /* Mux(v, x, x) == x */
1669 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1672 if (is_Cmp(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1673 ir_relation relation = get_Cmp_relation(sel);
1674 ir_node *f = get_Mux_false(n);
1675 ir_node *t = get_Mux_true(n);
1678 * Note further that these optimization work even for floating point
1679 * with NaN's because -NaN == NaN.
1680 * However, if +0 and -0 is handled differently, we cannot use the first one.
1682 ir_node *const cmp_l = get_Cmp_left(sel);
1683 ir_node *const cmp_r = get_Cmp_right(sel);
1686 case ir_relation_equal:
1687 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1688 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1690 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1695 case ir_relation_less_greater:
1696 case ir_relation_unordered_less_greater:
1697 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1698 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1700 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1709 * Note: normalization puts the constant on the right side,
1710 * so we check only one case.
1712 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1713 /* Mux(t CMP 0, X, t) */
1714 if (is_Minus(f) && get_Minus_op(f) == t) {
1715 /* Mux(t CMP 0, -t, t) */
1716 if (relation == ir_relation_equal) {
1717 /* Mux(t == 0, -t, t) ==> -t */
1719 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1720 } else if (relation == ir_relation_less_greater || relation == ir_relation_unordered_less_greater) {
1721 /* Mux(t != 0, -t, t) ==> t */
1723 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1733 * Remove Confirm nodes if setting is on.
1734 * Replace Confirms(x, '=', Constlike) by Constlike.
1736 static ir_node *equivalent_node_Confirm(ir_node *n)
1738 ir_node *pred = get_Confirm_value(n);
1739 ir_relation relation = get_Confirm_relation(n);
1741 while (is_Confirm(pred) && relation == get_Confirm_relation(pred)) {
1743 * rare case: two identical Confirms one after another,
1744 * replace the second one with the first.
1747 pred = get_Confirm_value(n);
1753 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1754 * perform no actual computation, as, e.g., the Id nodes. It does not create
1755 * new nodes. It is therefore safe to free n if the node returned is not n.
1756 * If a node returns a Tuple we can not just skip it. If the size of the
1757 * in array fits, we transform n into a tuple (e.g., Div).
1759 ir_node *equivalent_node(ir_node *n)
1761 if (n->op->ops.equivalent_node)
1762 return n->op->ops.equivalent_node(n);
1764 } /* equivalent_node */
1767 * Sets the default equivalent node operation for an ir_op_ops.
1769 * @param code the opcode for the default operation
1770 * @param ops the operations initialized
1775 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1779 ops->equivalent_node = equivalent_node_##a; \
1781 #define CASE_PROJ(a) \
1783 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1823 } /* firm_set_default_equivalent_node */
1826 * Returns non-zero if a node is a Phi node
1827 * with all predecessors constant.
1829 static int is_const_Phi(ir_node *n)
1833 if (! is_Phi(n) || get_irn_arity(n) == 0)
1835 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1836 if (! is_Const(get_irn_n(n, i)))
1840 } /* is_const_Phi */
1842 typedef ir_tarval *(*tarval_sub_type)(ir_tarval *a, ir_tarval *b, ir_mode *mode);
1843 typedef ir_tarval *(*tarval_binop_type)(ir_tarval *a, ir_tarval *b);
1846 * in reality eval_func should be tarval (*eval_func)() but incomplete
1847 * declarations are bad style and generate noisy warnings
1849 typedef void (*eval_func)(void);
1852 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1854 static ir_tarval *do_eval(eval_func eval, ir_tarval *a, ir_tarval *b, ir_mode *mode)
1856 if (eval == (eval_func) tarval_sub) {
1857 tarval_sub_type func = (tarval_sub_type)eval;
1859 return func(a, b, mode);
1861 tarval_binop_type func = (tarval_binop_type)eval;
1868 * Apply an evaluator on a binop with a constant operators (and one Phi).
1870 * @param phi the Phi node
1871 * @param other the other operand
1872 * @param eval an evaluator function
1873 * @param mode the mode of the result, may be different from the mode of the Phi!
1874 * @param left if non-zero, other is the left operand, else the right
1876 * @return a new Phi node if the conversion was successful, NULL else
1878 static ir_node *apply_binop_on_phi(ir_node *phi, ir_tarval *other, eval_func eval, ir_mode *mode, int left)
1884 int i, n = get_irn_arity(phi);
1886 NEW_ARR_A(void *, res, n);
1888 for (i = 0; i < n; ++i) {
1889 pred = get_irn_n(phi, i);
1890 tv = get_Const_tarval(pred);
1891 tv = do_eval(eval, other, tv, mode);
1893 if (tv == tarval_bad) {
1894 /* folding failed, bad */
1900 for (i = 0; i < n; ++i) {
1901 pred = get_irn_n(phi, i);
1902 tv = get_Const_tarval(pred);
1903 tv = do_eval(eval, tv, other, mode);
1905 if (tv == tarval_bad) {
1906 /* folding failed, bad */
1912 irg = get_irn_irg(phi);
1913 for (i = 0; i < n; ++i) {
1914 pred = get_irn_n(phi, i);
1915 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1917 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1918 } /* apply_binop_on_phi */
1921 * Apply an evaluator on a binop with two constant Phi.
1923 * @param a the left Phi node
1924 * @param b the right Phi node
1925 * @param eval an evaluator function
1926 * @param mode the mode of the result, may be different from the mode of the Phi!
1928 * @return a new Phi node if the conversion was successful, NULL else
1930 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, eval_func eval, ir_mode *mode)
1932 ir_tarval *tv_l, *tv_r, *tv;
1938 if (get_nodes_block(a) != get_nodes_block(b))
1941 n = get_irn_arity(a);
1942 NEW_ARR_A(void *, res, n);
1944 for (i = 0; i < n; ++i) {
1945 pred = get_irn_n(a, i);
1946 tv_l = get_Const_tarval(pred);
1947 pred = get_irn_n(b, i);
1948 tv_r = get_Const_tarval(pred);
1949 tv = do_eval(eval, tv_l, tv_r, mode);
1951 if (tv == tarval_bad) {
1952 /* folding failed, bad */
1957 irg = get_irn_irg(a);
1958 for (i = 0; i < n; ++i) {
1959 pred = get_irn_n(a, i);
1960 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1962 return new_r_Phi(get_nodes_block(a), n, (ir_node **)res, mode);
1963 } /* apply_binop_on_2_phis */
1966 * Apply an evaluator on a unop with a constant operator (a Phi).
1968 * @param phi the Phi node
1969 * @param eval an evaluator function
1971 * @return a new Phi node if the conversion was successful, NULL else
1973 static ir_node *apply_unop_on_phi(ir_node *phi, ir_tarval *(*eval)(ir_tarval *))
1980 int i, n = get_irn_arity(phi);
1982 NEW_ARR_A(void *, res, n);
1983 for (i = 0; i < n; ++i) {
1984 pred = get_irn_n(phi, i);
1985 tv = get_Const_tarval(pred);
1988 if (tv == tarval_bad) {
1989 /* folding failed, bad */
1994 mode = get_irn_mode(phi);
1995 irg = get_irn_irg(phi);
1996 for (i = 0; i < n; ++i) {
1997 pred = get_irn_n(phi, i);
1998 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
2000 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2001 } /* apply_unop_on_phi */
2004 * Apply a conversion on a constant operator (a Phi).
2006 * @param phi the Phi node
2008 * @return a new Phi node if the conversion was successful, NULL else
2010 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode)
2016 int i, n = get_irn_arity(phi);
2018 NEW_ARR_A(void *, res, n);
2019 for (i = 0; i < n; ++i) {
2020 pred = get_irn_n(phi, i);
2021 tv = get_Const_tarval(pred);
2022 tv = tarval_convert_to(tv, mode);
2024 if (tv == tarval_bad) {
2025 /* folding failed, bad */
2030 irg = get_irn_irg(phi);
2031 for (i = 0; i < n; ++i) {
2032 pred = get_irn_n(phi, i);
2033 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
2035 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2036 } /* apply_conv_on_phi */
2039 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
2040 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
2041 * If possible, remove the Conv's.
2043 static ir_node *transform_node_AddSub(ir_node *n)
2045 ir_mode *mode = get_irn_mode(n);
2047 if (mode_is_reference(mode)) {
2048 ir_node *left = get_binop_left(n);
2049 ir_node *right = get_binop_right(n);
2050 unsigned ref_bits = get_mode_size_bits(mode);
2052 if (is_Conv(left)) {
2053 ir_mode *lmode = get_irn_mode(left);
2054 unsigned bits = get_mode_size_bits(lmode);
2056 if (ref_bits == bits &&
2057 mode_is_int(lmode) &&
2058 get_mode_arithmetic(lmode) == irma_twos_complement) {
2059 ir_node *pre = get_Conv_op(left);
2060 ir_mode *pre_mode = get_irn_mode(pre);
2062 if (mode_is_int(pre_mode) &&
2063 get_mode_size_bits(pre_mode) == bits &&
2064 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2065 /* ok, this conv just changes to sign, moreover the calculation
2066 * is done with same number of bits as our address mode, so
2067 * we can ignore the conv as address calculation can be viewed
2068 * as either signed or unsigned
2070 set_binop_left(n, pre);
2075 if (is_Conv(right)) {
2076 ir_mode *rmode = get_irn_mode(right);
2077 unsigned bits = get_mode_size_bits(rmode);
2079 if (ref_bits == bits &&
2080 mode_is_int(rmode) &&
2081 get_mode_arithmetic(rmode) == irma_twos_complement) {
2082 ir_node *pre = get_Conv_op(right);
2083 ir_mode *pre_mode = get_irn_mode(pre);
2085 if (mode_is_int(pre_mode) &&
2086 get_mode_size_bits(pre_mode) == bits &&
2087 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2088 /* ok, this conv just changes to sign, moreover the calculation
2089 * is done with same number of bits as our address mode, so
2090 * we can ignore the conv as address calculation can be viewed
2091 * as either signed or unsigned
2093 set_binop_right(n, pre);
2098 /* let address arithmetic use unsigned modes */
2099 if (is_Const(right)) {
2100 ir_mode *rmode = get_irn_mode(right);
2102 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2103 /* convert a AddP(P, *s) into AddP(P, *u) */
2104 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2106 ir_node *pre = new_r_Conv(get_nodes_block(n), right, nm);
2107 set_binop_right(n, pre);
2113 } /* transform_node_AddSub */
2115 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2118 if (is_Const(b) && is_const_Phi(a)) { \
2119 /* check for Op(Phi, Const) */ \
2120 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2122 else if (is_Const(a) && is_const_Phi(b)) { \
2123 /* check for Op(Const, Phi) */ \
2124 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2126 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2127 /* check for Op(Phi, Phi) */ \
2128 c = apply_binop_on_2_phis(a, b, eval, mode); \
2131 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2136 #define HANDLE_UNOP_PHI(eval, a, c) \
2139 if (is_const_Phi(a)) { \
2140 /* check for Op(Phi) */ \
2141 c = apply_unop_on_phi(a, eval); \
2143 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2150 * Do the AddSub optimization, then Transform
2151 * Constant folding on Phi
2152 * Add(a,a) -> Mul(a, 2)
2153 * Add(Mul(a, x), a) -> Mul(a, x+1)
2154 * if the mode is integer or float.
2155 * Transform Add(a,-b) into Sub(a,b).
2156 * Reassociation might fold this further.
2158 static ir_node *transform_node_Add(ir_node *n)
2161 ir_node *a, *b, *c, *oldn = n;
2162 vrp_attr *a_vrp, *b_vrp;
2164 n = transform_node_AddSub(n);
2166 a = get_Add_left(n);
2167 b = get_Add_right(n);
2169 mode = get_irn_mode(n);
2171 if (mode_is_reference(mode)) {
2172 ir_mode *lmode = get_irn_mode(a);
2174 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2175 /* an Add(a, NULL) is a hidden Conv */
2176 dbg_info *dbg = get_irn_dbg_info(n);
2177 return new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2181 HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode);
2183 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2184 if (mode_is_float(mode)) {
2185 ir_graph *irg = get_irn_irg(n);
2186 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2190 if (mode_is_num(mode)) {
2191 ir_graph *irg = get_irn_irg(n);
2192 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2193 if (!is_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP)
2194 && a == b && mode_is_int(mode)) {
2195 ir_node *block = get_nodes_block(n);
2198 get_irn_dbg_info(n),
2201 new_r_Const_long(irg, mode, 2),
2203 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2208 get_irn_dbg_info(n),
2213 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2218 get_irn_dbg_info(n),
2223 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2226 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2227 /* Here we rely on constants be on the RIGHT side */
2229 ir_node *op = get_Not_op(a);
2231 if (is_Const(b) && is_Const_one(b)) {
2233 ir_node *blk = get_nodes_block(n);
2234 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, mode);
2235 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2240 n = new_r_Const(irg, get_mode_minus_one(mode));
2241 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2246 ir_node *op = get_Not_op(b);
2250 n = new_r_Const(irg, get_mode_minus_one(mode));
2251 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2258 a_vrp = vrp_get_info(a);
2259 b_vrp = vrp_get_info(b);
2261 if (a_vrp && b_vrp) {
2262 ir_tarval *c = tarval_and(a_vrp->bits_not_set, b_vrp->bits_not_set);
2264 if (tarval_is_null(c)) {
2265 dbg_info *dbgi = get_irn_dbg_info(n);
2266 return new_rd_Or(dbgi, get_nodes_block(n), a, b, mode);
2270 } /* transform_node_Add */
2273 * returns -cnst or NULL if impossible
2275 static ir_node *const_negate(ir_node *cnst)
2277 ir_tarval *tv = tarval_neg(get_Const_tarval(cnst));
2278 dbg_info *dbgi = get_irn_dbg_info(cnst);
2279 ir_graph *irg = get_irn_irg(cnst);
2280 if (tv == tarval_bad) return NULL;
2281 return new_rd_Const(dbgi, irg, tv);
2285 * Do the AddSub optimization, then Transform
2286 * Constant folding on Phi
2287 * Sub(0,a) -> Minus(a)
2288 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2289 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2290 * Sub(Add(a, x), x) -> a
2291 * Sub(x, Add(x, a)) -> -a
2292 * Sub(x, Const) -> Add(x, -Const)
2294 static ir_node *transform_node_Sub(ir_node *n)
2300 n = transform_node_AddSub(n);
2302 a = get_Sub_left(n);
2303 b = get_Sub_right(n);
2305 mode = get_irn_mode(n);
2307 if (mode_is_int(mode)) {
2308 ir_mode *lmode = get_irn_mode(a);
2310 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2311 /* a Sub(a, NULL) is a hidden Conv */
2312 dbg_info *dbg = get_irn_dbg_info(n);
2313 n = new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2314 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2318 if (mode == lmode &&
2319 get_mode_arithmetic(mode) == irma_twos_complement &&
2321 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2323 dbg_info *dbg = get_irn_dbg_info(n);
2324 n = new_rd_Not(dbg, get_nodes_block(n), b, mode);
2325 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2331 HANDLE_BINOP_PHI((eval_func) tarval_sub, a, b, c, mode);
2333 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2334 if (mode_is_float(mode)) {
2335 ir_graph *irg = get_irn_irg(n);
2336 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2340 if (is_Const(b) && !mode_is_reference(get_irn_mode(b))) {
2341 /* a - C -> a + (-C) */
2342 ir_node *cnst = const_negate(b);
2344 ir_node *block = get_nodes_block(n);
2345 dbg_info *dbgi = get_irn_dbg_info(n);
2347 n = new_rd_Add(dbgi, block, a, cnst, mode);
2348 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2353 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2354 dbg_info *dbg = get_irn_dbg_info(n);
2355 ir_node *block = get_nodes_block(n);
2356 ir_node *left = get_Minus_op(a);
2357 ir_node *add = new_rd_Add(dbg, block, left, b, mode);
2359 n = new_rd_Minus(dbg, block, add, mode);
2360 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2362 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2363 dbg_info *dbg = get_irn_dbg_info(n);
2364 ir_node *block = get_nodes_block(n);
2365 ir_node *right = get_Minus_op(b);
2367 n = new_rd_Add(dbg, block, a, right, mode);
2368 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2370 } else if (is_Sub(b)) {
2371 /* a - (b - c) -> a + (c - b)
2372 * -> (a - b) + c iff (b - c) is a pointer */
2373 dbg_info *s_dbg = get_irn_dbg_info(b);
2374 ir_node *s_left = get_Sub_left(b);
2375 ir_node *s_right = get_Sub_right(b);
2376 ir_mode *s_mode = get_irn_mode(b);
2377 if (mode_is_reference(s_mode)) {
2378 ir_node *lowest_block = get_nodes_block(n); /* a and b are live here */
2379 ir_node *sub = new_rd_Sub(s_dbg, lowest_block, a, s_left, mode);
2380 dbg_info *a_dbg = get_irn_dbg_info(n);
2383 s_right = new_r_Conv(lowest_block, s_right, mode);
2384 n = new_rd_Add(a_dbg, lowest_block, sub, s_right, mode);
2386 ir_node *s_block = get_nodes_block(b);
2387 ir_node *sub = new_rd_Sub(s_dbg, s_block, s_right, s_left, s_mode);
2388 dbg_info *a_dbg = get_irn_dbg_info(n);
2389 ir_node *a_block = get_nodes_block(n);
2391 n = new_rd_Add(a_dbg, a_block, a, sub, mode);
2393 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2395 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2396 ir_node *m_right = get_Mul_right(b);
2397 if (is_Const(m_right)) {
2398 ir_node *cnst2 = const_negate(m_right);
2399 if (cnst2 != NULL) {
2400 dbg_info *m_dbg = get_irn_dbg_info(b);
2401 ir_node *m_block = get_nodes_block(b);
2402 ir_node *m_left = get_Mul_left(b);
2403 ir_mode *m_mode = get_irn_mode(b);
2404 ir_node *mul = new_rd_Mul(m_dbg, m_block, m_left, cnst2, m_mode);
2405 dbg_info *a_dbg = get_irn_dbg_info(n);
2406 ir_node *a_block = get_nodes_block(n);
2408 n = new_rd_Add(a_dbg, a_block, a, mul, mode);
2409 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2415 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2416 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2418 get_irn_dbg_info(n),
2422 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2426 if (mode_wrap_around(mode)) {
2427 ir_node *left = get_Add_left(a);
2428 ir_node *right = get_Add_right(a);
2430 /* FIXME: Does the Conv's work only for two complement or generally? */
2432 if (mode != get_irn_mode(right)) {
2433 /* This Sub is an effective Cast */
2434 right = new_r_Conv(get_nodes_block(n), right, mode);
2437 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2439 } else if (right == b) {
2440 if (mode != get_irn_mode(left)) {
2441 /* This Sub is an effective Cast */
2442 left = new_r_Conv(get_nodes_block(n), left, mode);
2445 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2451 if (mode_wrap_around(mode)) {
2452 ir_node *left = get_Add_left(b);
2453 ir_node *right = get_Add_right(b);
2455 /* FIXME: Does the Conv's work only for two complement or generally? */
2457 ir_mode *r_mode = get_irn_mode(right);
2459 n = new_r_Minus(get_nodes_block(n), right, r_mode);
2460 if (mode != r_mode) {
2461 /* This Sub is an effective Cast */
2462 n = new_r_Conv(get_nodes_block(n), n, mode);
2464 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2466 } else if (right == a) {
2467 ir_mode *l_mode = get_irn_mode(left);
2469 n = new_r_Minus(get_nodes_block(n), left, l_mode);
2470 if (mode != l_mode) {
2471 /* This Sub is an effective Cast */
2472 n = new_r_Conv(get_nodes_block(n), n, mode);
2474 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2479 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2480 ir_mode *mode = get_irn_mode(a);
2482 if (mode == get_irn_mode(b)) {
2484 ir_node *op_a = get_Conv_op(a);
2485 ir_node *op_b = get_Conv_op(b);
2487 /* check if it's allowed to skip the conv */
2488 ma = get_irn_mode(op_a);
2489 mb = get_irn_mode(op_b);
2491 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2492 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2495 set_Sub_right(n, b);
2501 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2502 if (!is_reassoc_running() && is_Mul(a)) {
2503 ir_node *ma = get_Mul_left(a);
2504 ir_node *mb = get_Mul_right(a);
2507 ir_node *blk = get_nodes_block(n);
2508 ir_graph *irg = get_irn_irg(n);
2510 get_irn_dbg_info(n),
2514 get_irn_dbg_info(n),
2517 new_r_Const(irg, get_mode_one(mode)),
2520 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2522 } else if (mb == b) {
2523 ir_node *blk = get_nodes_block(n);
2524 ir_graph *irg = get_irn_irg(n);
2526 get_irn_dbg_info(n),
2530 get_irn_dbg_info(n),
2533 new_r_Const(irg, get_mode_one(mode)),
2536 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2540 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2541 ir_node *x = get_Sub_left(a);
2542 ir_node *y = get_Sub_right(a);
2543 ir_node *blk = get_nodes_block(n);
2544 ir_mode *m_b = get_irn_mode(b);
2545 ir_mode *m_y = get_irn_mode(y);
2549 /* Determine the right mode for the Add. */
2552 else if (mode_is_reference(m_b))
2554 else if (mode_is_reference(m_y))
2558 * Both modes are different but none is reference,
2559 * happens for instance in SubP(SubP(P, Iu), Is).
2560 * We have two possibilities here: Cast or ignore.
2561 * Currently we ignore this case.
2566 add = new_r_Add(blk, y, b, add_mode);
2568 n = new_rd_Sub(get_irn_dbg_info(n), blk, x, add, mode);
2569 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2573 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2574 if (is_Const(a) && is_Not(b)) {
2575 /* c - ~X = X + (c+1) */
2576 ir_tarval *tv = get_Const_tarval(a);
2578 tv = tarval_add(tv, get_mode_one(mode));
2579 if (tv != tarval_bad) {
2580 ir_node *blk = get_nodes_block(n);
2581 ir_graph *irg = get_irn_irg(n);
2582 ir_node *c = new_r_Const(irg, tv);
2583 n = new_rd_Add(get_irn_dbg_info(n), blk, get_Not_op(b), c, mode);
2584 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2590 } /* transform_node_Sub */
2593 * Several transformation done on n*n=2n bits mul.
2594 * These transformations must be done here because new nodes may be produced.
2596 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode)
2599 ir_node *a = get_Mul_left(n);
2600 ir_node *b = get_Mul_right(n);
2601 ir_tarval *ta = value_of(a);
2602 ir_tarval *tb = value_of(b);
2603 ir_mode *smode = get_irn_mode(a);
2605 if (ta == get_mode_one(smode)) {
2606 /* (L)1 * (L)b = (L)b */
2607 ir_node *blk = get_nodes_block(n);
2608 n = new_rd_Conv(get_irn_dbg_info(n), blk, b, mode);
2609 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2612 else if (ta == get_mode_minus_one(smode)) {
2613 /* (L)-1 * (L)b = (L)b */
2614 ir_node *blk = get_nodes_block(n);
2615 n = new_rd_Minus(get_irn_dbg_info(n), blk, b, smode);
2616 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2617 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2620 if (tb == get_mode_one(smode)) {
2621 /* (L)a * (L)1 = (L)a */
2622 ir_node *blk = get_irn_n(a, -1);
2623 n = new_rd_Conv(get_irn_dbg_info(n), blk, a, mode);
2624 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2627 else if (tb == get_mode_minus_one(smode)) {
2628 /* (L)a * (L)-1 = (L)-a */
2629 ir_node *blk = get_nodes_block(n);
2630 n = new_rd_Minus(get_irn_dbg_info(n), blk, a, smode);
2631 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2632 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2639 * Transform Mul(a,-1) into -a.
2640 * Do constant evaluation of Phi nodes.
2641 * Do architecture dependent optimizations on Mul nodes
2643 static ir_node *transform_node_Mul(ir_node *n)
2645 ir_node *c, *oldn = n;
2646 ir_mode *mode = get_irn_mode(n);
2647 ir_node *a = get_Mul_left(n);
2648 ir_node *b = get_Mul_right(n);
2650 if (is_Bad(a) || is_Bad(b))
2653 if (mode != get_irn_mode(a))
2654 return transform_node_Mul2n(n, mode);
2656 HANDLE_BINOP_PHI((eval_func) tarval_mul, a, b, c, mode);
2658 if (mode_is_signed(mode)) {
2661 if (value_of(a) == get_mode_minus_one(mode))
2663 else if (value_of(b) == get_mode_minus_one(mode))
2666 n = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), r, mode);
2667 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2672 if (is_Const(b)) { /* (-a) * const -> a * -const */
2673 ir_node *cnst = const_negate(b);
2675 dbg_info *dbgi = get_irn_dbg_info(n);
2676 ir_node *block = get_nodes_block(n);
2677 n = new_rd_Mul(dbgi, block, get_Minus_op(a), cnst, mode);
2678 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2681 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2682 dbg_info *dbgi = get_irn_dbg_info(n);
2683 ir_node *block = get_nodes_block(n);
2684 n = new_rd_Mul(dbgi, block, get_Minus_op(a), get_Minus_op(b), mode);
2685 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2687 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2688 ir_node *sub_l = get_Sub_left(b);
2689 ir_node *sub_r = get_Sub_right(b);
2690 dbg_info *dbgi = get_irn_dbg_info(n);
2691 ir_node *block = get_nodes_block(n);
2692 ir_node *new_b = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2693 n = new_rd_Mul(dbgi, block, get_Minus_op(a), new_b, mode);
2694 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2697 } else if (is_Minus(b)) {
2698 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2699 ir_node *sub_l = get_Sub_left(a);
2700 ir_node *sub_r = get_Sub_right(a);
2701 dbg_info *dbgi = get_irn_dbg_info(n);
2702 ir_node *block = get_nodes_block(n);
2703 ir_node *new_a = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2704 n = new_rd_Mul(dbgi, block, new_a, get_Minus_op(b), mode);
2705 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2708 } else if (is_Shl(a)) {
2709 ir_node *const shl_l = get_Shl_left(a);
2710 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2711 /* (1 << x) * b -> b << x */
2712 dbg_info *const dbgi = get_irn_dbg_info(n);
2713 ir_node *const block = get_nodes_block(n);
2714 ir_node *const shl_r = get_Shl_right(a);
2715 n = new_rd_Shl(dbgi, block, b, shl_r, mode);
2716 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2719 } else if (is_Shl(b)) {
2720 ir_node *const shl_l = get_Shl_left(b);
2721 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2722 /* a * (1 << x) -> a << x */
2723 dbg_info *const dbgi = get_irn_dbg_info(n);
2724 ir_node *const block = get_nodes_block(n);
2725 ir_node *const shl_r = get_Shl_right(b);
2726 n = new_rd_Shl(dbgi, block, a, shl_r, mode);
2727 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2731 if (get_mode_arithmetic(mode) == irma_ieee754) {
2733 ir_tarval *tv = get_Const_tarval(a);
2734 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2735 && !tarval_is_negative(tv)) {
2736 /* 2.0 * b = b + b */
2737 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), b, b, mode);
2738 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2742 else if (is_Const(b)) {
2743 ir_tarval *tv = get_Const_tarval(b);
2744 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2745 && !tarval_is_negative(tv)) {
2746 /* a * 2.0 = a + a */
2747 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), a, a, mode);
2748 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2753 return arch_dep_replace_mul_with_shifts(n);
2754 } /* transform_node_Mul */
2757 * Transform a Div Node.
2759 static ir_node *transform_node_Div(ir_node *n)
2761 ir_mode *mode = get_Div_resmode(n);
2762 ir_node *a = get_Div_left(n);
2763 ir_node *b = get_Div_right(n);
2765 const ir_node *dummy;
2767 if (mode_is_int(mode)) {
2768 if (is_Const(b) && is_const_Phi(a)) {
2769 /* check for Div(Phi, Const) */
2770 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0);
2772 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2775 } else if (is_Const(a) && is_const_Phi(b)) {
2776 /* check for Div(Const, Phi) */
2777 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1);
2779 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2782 } else if (is_const_Phi(a) && is_const_Phi(b)) {
2783 /* check for Div(Phi, Phi) */
2784 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode);
2786 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2791 if (a == b && value_not_zero(a, &dummy)) {
2792 ir_graph *irg = get_irn_irg(n);
2793 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2794 value = new_r_Const(irg, get_mode_one(mode));
2795 DBG_OPT_CSTEVAL(n, value);
2798 if (mode_is_signed(mode) && is_Const(b)) {
2799 ir_tarval *tv = get_Const_tarval(b);
2801 if (tv == get_mode_minus_one(mode)) {
2803 value = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
2804 DBG_OPT_CSTEVAL(n, value);
2808 /* Try architecture dependent optimization */
2809 value = arch_dep_replace_div_by_const(n);
2812 assert(mode_is_float(mode));
2814 /* Optimize x/c to x*(1/c) */
2815 if (get_mode_arithmetic(mode) == irma_ieee754) {
2816 ir_tarval *tv = value_of(b);
2818 if (tv != tarval_bad) {
2819 int rem = tarval_fp_ops_enabled();
2822 * Floating point constant folding might be disabled here to
2824 * However, as we check for exact result, doing it is safe.
2827 tarval_enable_fp_ops(1);
2828 tv = tarval_div(get_mode_one(mode), tv);
2829 tarval_enable_fp_ops(rem);
2831 /* Do the transformation if the result is either exact or we are
2832 not using strict rules. */
2833 if (tv != tarval_bad &&
2834 (tarval_ieee754_get_exact() || (get_irg_fp_model(get_irn_irg(n)) & fp_strict_algebraic) == 0)) {
2835 ir_node *block = get_nodes_block(n);
2836 ir_graph *irg = get_irn_irg(block);
2837 ir_node *c = new_r_Const(irg, tv);
2838 dbg_info *dbgi = get_irn_dbg_info(n);
2839 value = new_rd_Mul(dbgi, block, a, c, mode);
2852 /* Turn Div into a tuple (mem, jmp, bad, value) */
2853 mem = get_Div_mem(n);
2854 blk = get_nodes_block(n);
2855 irg = get_irn_irg(blk);
2857 /* skip a potential Pin */
2858 mem = skip_Pin(mem);
2859 turn_into_tuple(n, pn_Div_max);
2860 set_Tuple_pred(n, pn_Div_M, mem);
2861 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(blk));
2862 set_Tuple_pred(n, pn_Div_X_except, get_irg_bad(irg));
2863 set_Tuple_pred(n, pn_Div_res, value);
2866 } /* transform_node_Div */
2869 * Transform a Mod node.
2871 static ir_node *transform_node_Mod(ir_node *n)
2873 ir_mode *mode = get_Mod_resmode(n);
2874 ir_node *a = get_Mod_left(n);
2875 ir_node *b = get_Mod_right(n);
2880 if (is_Const(b) && is_const_Phi(a)) {
2881 /* check for Div(Phi, Const) */
2882 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0);
2884 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2888 else if (is_Const(a) && is_const_Phi(b)) {
2889 /* check for Div(Const, Phi) */
2890 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1);
2892 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2896 else if (is_const_Phi(a) && is_const_Phi(b)) {
2897 /* check for Div(Phi, Phi) */
2898 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode);
2900 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2907 irg = get_irn_irg(n);
2908 if (tv != tarval_bad) {
2909 value = new_r_Const(irg, tv);
2911 DBG_OPT_CSTEVAL(n, value);
2914 ir_node *a = get_Mod_left(n);
2915 ir_node *b = get_Mod_right(n);
2916 const ir_node *dummy;
2918 if (a == b && value_not_zero(a, &dummy)) {
2919 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2920 value = new_r_Const(irg, get_mode_null(mode));
2921 DBG_OPT_CSTEVAL(n, value);
2924 if (mode_is_signed(mode) && is_Const(b)) {
2925 ir_tarval *tv = get_Const_tarval(b);
2927 if (tv == get_mode_minus_one(mode)) {
2929 value = new_r_Const(irg, get_mode_null(mode));
2930 DBG_OPT_CSTEVAL(n, value);
2934 /* Try architecture dependent optimization */
2935 value = arch_dep_replace_mod_by_const(n);
2944 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2945 mem = get_Mod_mem(n);
2946 blk = get_nodes_block(n);
2947 irg = get_irn_irg(blk);
2949 /* skip a potential Pin */
2950 mem = skip_Pin(mem);
2951 turn_into_tuple(n, pn_Mod_max);
2952 set_Tuple_pred(n, pn_Mod_M, mem);
2953 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(blk));
2954 set_Tuple_pred(n, pn_Mod_X_except, get_irg_bad(irg));
2955 set_Tuple_pred(n, pn_Mod_res, value);
2958 } /* transform_node_Mod */
2961 * Transform a Cond node.
2963 * Replace the Cond by a Jmp if it branches on a constant
2966 static ir_node *transform_node_Cond(ir_node *n)
2969 ir_node *a = get_Cond_selector(n);
2970 ir_tarval *ta = value_of(a);
2971 ir_graph *irg = get_irn_irg(n);
2974 /* we need block info which is not available in floating irgs */
2975 if (get_irg_pinned(irg) == op_pin_state_floats)
2978 if ((ta != tarval_bad) &&
2979 (get_irn_mode(a) == mode_b) &&
2980 (get_opt_unreachable_code())) {
2981 /* It's a boolean Cond, branching on a boolean constant.
2982 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2983 ir_node *blk = get_nodes_block(n);
2984 jmp = new_r_Jmp(blk);
2985 turn_into_tuple(n, pn_Cond_max);
2986 if (ta == tarval_b_true) {
2987 set_Tuple_pred(n, pn_Cond_false, get_irg_bad(irg));
2988 set_Tuple_pred(n, pn_Cond_true, jmp);
2990 set_Tuple_pred(n, pn_Cond_false, jmp);
2991 set_Tuple_pred(n, pn_Cond_true, get_irg_bad(irg));
2993 /* We might generate an endless loop, so keep it alive. */
2994 add_End_keepalive(get_irg_end(irg), blk);
2997 } /* transform_node_Cond */
3000 * Prototype of a recursive transform function
3001 * for bitwise distributive transformations.
3003 typedef ir_node* (*recursive_transform)(ir_node *n);
3006 * makes use of distributive laws for and, or, eor
3007 * and(a OP c, b OP c) -> and(a, b) OP c
3008 * note, might return a different op than n
3010 static ir_node *transform_bitwise_distributive(ir_node *n,
3011 recursive_transform trans_func)
3014 ir_node *a = get_binop_left(n);
3015 ir_node *b = get_binop_right(n);
3016 ir_op *op = get_irn_op(a);
3017 ir_op *op_root = get_irn_op(n);
3019 if (op != get_irn_op(b))
3022 /* and(conv(a), conv(b)) -> conv(and(a,b)) */
3023 if (op == op_Conv) {
3024 ir_node *a_op = get_Conv_op(a);
3025 ir_node *b_op = get_Conv_op(b);
3026 ir_mode *a_mode = get_irn_mode(a_op);
3027 ir_mode *b_mode = get_irn_mode(b_op);
3028 if (a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3029 ir_node *blk = get_nodes_block(n);
3032 set_binop_left(n, a_op);
3033 set_binop_right(n, b_op);
3034 set_irn_mode(n, a_mode);
3036 n = new_r_Conv(blk, n, get_irn_mode(oldn));
3038 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
3044 /* nothing to gain here */
3048 if (op == op_Shrs || op == op_Shr || op == op_Shl
3049 || op == op_And || op == op_Or || op == op_Eor) {
3050 ir_node *a_left = get_binop_left(a);
3051 ir_node *a_right = get_binop_right(a);
3052 ir_node *b_left = get_binop_left(b);
3053 ir_node *b_right = get_binop_right(b);
3055 ir_node *op1 = NULL;
3056 ir_node *op2 = NULL;
3058 if (is_op_commutative(op)) {
3059 if (a_left == b_left) {
3063 } else if (a_left == b_right) {
3067 } else if (a_right == b_left) {
3073 if (a_right == b_right) {
3080 /* (a sop c) & (b sop c) => (a & b) sop c */
3081 ir_node *blk = get_nodes_block(n);
3083 ir_node *new_n = exact_copy(n);
3084 set_binop_left(new_n, op1);
3085 set_binop_right(new_n, op2);
3086 new_n = trans_func(new_n);
3088 if (op_root == op_Eor && op == op_Or) {
3089 dbg_info *dbgi = get_irn_dbg_info(n);
3090 ir_mode *mode = get_irn_mode(c);
3092 c = new_rd_Not(dbgi, blk, c, mode);
3093 n = new_rd_And(dbgi, blk, new_n, c, mode);
3096 set_nodes_block(n, blk);
3097 set_binop_left(n, new_n);
3098 set_binop_right(n, c);
3102 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3111 * Create a 0 constant of given mode.
3113 static ir_node *create_zero_const(ir_graph *irg, ir_mode *mode)
3115 ir_tarval *tv = get_mode_null(mode);
3116 ir_node *cnst = new_r_Const(irg, tv);
3124 static ir_node *transform_node_And(ir_node *n)
3126 ir_node *c, *oldn = n;
3127 ir_node *a = get_And_left(n);
3128 ir_node *b = get_And_right(n);
3130 vrp_attr *a_vrp, *b_vrp;
3132 if (is_Cmp(a) && is_Cmp(b)) {
3133 ir_node *a_left = get_Cmp_left(a);
3134 ir_node *a_right = get_Cmp_right(a);
3135 ir_node *b_left = get_Cmp_left(b);
3136 ir_node *b_right = get_Cmp_right(b);
3137 ir_relation a_relation = get_Cmp_relation(a);
3138 ir_relation b_relation = get_Cmp_relation(b);
3139 /* we can combine the relations of two compares with the same
3141 if (a_left == b_left && b_left == b_right) {
3142 dbg_info *dbgi = get_irn_dbg_info(n);
3143 ir_node *block = get_nodes_block(n);
3144 ir_relation new_relation = a_relation & b_relation;
3145 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
3147 /* Cmp(a==0) and Cmp(b==0) can be optimized to Cmp(a|b==0) */
3148 if (is_Const(a_right) && is_Const_null(a_right)
3149 && is_Const(b_right) && is_Const_null(b_right)
3150 && a_relation == b_relation && a_relation == ir_relation_equal
3151 && !mode_is_float(get_irn_mode(a_left))
3152 && !mode_is_float(get_irn_mode(b_left))) {
3153 dbg_info *dbgi = get_irn_dbg_info(n);
3154 ir_node *block = get_nodes_block(n);
3155 ir_mode *mode = get_irn_mode(a_left);
3156 ir_node *n_b_left = get_irn_mode(b_left) != mode ?
3157 new_rd_Conv(dbgi, block, b_left, mode) : b_left;
3158 ir_node *or = new_rd_Or(dbgi, block, a_left, n_b_left, mode);
3159 ir_graph *irg = get_irn_irg(n);
3160 ir_node *zero = create_zero_const(irg, mode);
3161 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_equal);
3165 mode = get_irn_mode(n);
3166 HANDLE_BINOP_PHI((eval_func) tarval_and, a, b, c, mode);
3170 ir_node *op = get_Not_op(b);
3172 ir_node *ba = get_And_left(op);
3173 ir_node *bb = get_And_right(op);
3175 /* it's enough to test the following cases due to normalization! */
3176 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3177 /* (a|b) & ~(a&b) = a^b */
3178 ir_node *block = get_nodes_block(n);
3180 n = new_rd_Eor(get_irn_dbg_info(n), block, ba, bb, mode);
3181 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3189 ir_node *op = get_Not_op(a);
3191 ir_node *aa = get_And_left(op);
3192 ir_node *ab = get_And_right(op);
3194 /* it's enough to test the following cases due to normalization! */
3195 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3196 /* (a|b) & ~(a&b) = a^b */
3197 ir_node *block = get_nodes_block(n);
3199 n = new_rd_Eor(get_irn_dbg_info(n), block, aa, ab, mode);
3200 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3207 ir_node *al = get_Eor_left(a);
3208 ir_node *ar = get_Eor_right(a);
3211 /* (b ^ a) & b -> ~a & b */
3212 dbg_info *dbg = get_irn_dbg_info(n);
3213 ir_node *block = get_nodes_block(n);
3215 ar = new_rd_Not(dbg, block, ar, mode);
3216 n = new_rd_And(dbg, block, ar, b, mode);
3217 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3221 /* (a ^ b) & b -> ~a & b */
3222 dbg_info *dbg = get_irn_dbg_info(n);
3223 ir_node *block = get_nodes_block(n);
3225 al = new_rd_Not(dbg, block, al, mode);
3226 n = new_rd_And(dbg, block, al, b, mode);
3227 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3232 ir_node *bl = get_Eor_left(b);
3233 ir_node *br = get_Eor_right(b);
3236 /* a & (a ^ b) -> a & ~b */
3237 dbg_info *dbg = get_irn_dbg_info(n);
3238 ir_node *block = get_nodes_block(n);
3240 br = new_rd_Not(dbg, block, br, mode);
3241 n = new_rd_And(dbg, block, br, a, mode);
3242 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3246 /* a & (b ^ a) -> a & ~b */
3247 dbg_info *dbg = get_irn_dbg_info(n);
3248 ir_node *block = get_nodes_block(n);
3250 bl = new_rd_Not(dbg, block, bl, mode);
3251 n = new_rd_And(dbg, block, bl, a, mode);
3252 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3256 if (is_Not(a) && is_Not(b)) {
3257 /* ~a & ~b = ~(a|b) */
3258 ir_node *block = get_nodes_block(n);
3259 ir_mode *mode = get_irn_mode(n);
3263 n = new_rd_Or(get_irn_dbg_info(n), block, a, b, mode);
3264 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
3265 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3269 b_vrp = vrp_get_info(b);
3270 if (is_Const(a) && b_vrp && (tarval_cmp(tarval_or(get_Const_tarval(a),
3271 b_vrp->bits_not_set), get_Const_tarval(a)) == ir_relation_equal)) {
3277 a_vrp = vrp_get_info(a);
3278 if (is_Const(b) && a_vrp && (tarval_cmp(tarval_or(get_Const_tarval(b),
3279 a_vrp->bits_not_set), get_Const_tarval(b)) == ir_relation_equal)) {
3283 n = transform_bitwise_distributive(n, transform_node_And);
3286 } /* transform_node_And */
3288 /* the order of the values is important! */
3289 typedef enum const_class {
3295 static const_class classify_const(const ir_node* n)
3297 if (is_Const(n)) return const_const;
3298 if (is_irn_constlike(n)) return const_like;
3303 * Determines whether r is more constlike or has a larger index (in that order)
3306 static bool operands_are_normalized(const ir_node *l, const ir_node *r)
3308 const const_class l_order = classify_const(l);
3309 const const_class r_order = classify_const(r);
3311 l_order > r_order ||
3312 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3318 static ir_node *transform_node_Eor(ir_node *n)
3320 ir_node *c, *oldn = n;
3321 ir_node *a = get_Eor_left(n);
3322 ir_node *b = get_Eor_right(n);
3323 ir_mode *mode = get_irn_mode(n);
3325 /* we can combine the relations of two compares with the same operands */
3326 if (is_Cmp(a) && is_Cmp(b)) {
3327 ir_node *a_left = get_Cmp_left(a);
3328 ir_node *a_right = get_Cmp_left(a);
3329 ir_node *b_left = get_Cmp_left(b);
3330 ir_node *b_right = get_Cmp_right(b);
3331 if (a_left == b_left && b_left == b_right) {
3332 dbg_info *dbgi = get_irn_dbg_info(n);
3333 ir_node *block = get_nodes_block(n);
3334 ir_relation a_relation = get_Cmp_relation(a);
3335 ir_relation b_relation = get_Cmp_relation(b);
3336 ir_relation new_relation = a_relation ^ b_relation;
3337 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
3341 HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode);
3343 /* normalize not nodes... ~a ^ b <=> a ^ ~b */
3344 if (is_Not(a) && operands_are_normalized(get_Not_op(a), b)) {
3345 dbg_info *dbg = get_irn_dbg_info(n);
3346 ir_node *block = get_nodes_block(n);
3347 ir_node *new_not = new_rd_Not(dbg, block, b, mode);
3348 ir_node *new_left = get_Not_op(a);
3349 n = new_rd_Eor(dbg, block, new_left, new_not, mode);
3350 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3352 } else if (is_Not(b) && !operands_are_normalized(a, get_Not_op(b))) {
3353 dbg_info *dbg = get_irn_dbg_info(n);
3354 ir_node *block = get_nodes_block(n);
3355 ir_node *new_not = new_rd_Not(dbg, block, a, mode);
3356 ir_node *new_right = get_Not_op(b);
3357 n = new_rd_Eor(dbg, block, new_not, new_right, mode);
3358 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3362 /* x ^ 1...1 -> ~1 */
3363 if (is_Const(b) && is_Const_all_one(b)) {
3364 n = new_r_Not(get_nodes_block(n), a, mode);
3365 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3369 n = transform_bitwise_distributive(n, transform_node_Eor);
3371 } /* transform_node_Eor */
3376 static ir_node *transform_node_Not(ir_node *n)
3378 ir_node *c, *oldn = n;
3379 ir_node *a = get_Not_op(n);
3380 ir_mode *mode = get_irn_mode(n);
3382 HANDLE_UNOP_PHI(tarval_not,a,c);
3384 /* check for a boolean Not */
3386 dbg_info *dbgi = get_irn_dbg_info(a);
3387 ir_node *block = get_nodes_block(a);
3388 ir_relation relation = get_Cmp_relation(a);
3389 relation = get_negated_relation(relation);
3390 n = new_rd_Cmp(dbgi, block, get_Cmp_left(a), get_Cmp_right(a), relation);
3391 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3395 /* normalize ~(a ^ b) => a ^ ~b */
3397 dbg_info *dbg = get_irn_dbg_info(n);
3398 ir_node *block = get_nodes_block(n);
3399 ir_node *eor_right = get_Eor_right(a);
3400 ir_node *eor_left = get_Eor_left(a);
3401 eor_right = new_rd_Not(dbg, block, eor_right, mode);
3402 n = new_rd_Eor(dbg, block, eor_left, eor_right, mode);
3406 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3407 if (is_Minus(a)) { /* ~-x -> x + -1 */
3408 dbg_info *dbg = get_irn_dbg_info(n);
3409 ir_graph *irg = get_irn_irg(n);
3410 ir_node *block = get_nodes_block(n);
3411 ir_node *add_l = get_Minus_op(a);
3412 ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode));
3413 n = new_rd_Add(dbg, block, add_l, add_r, mode);
3414 } else if (is_Add(a)) {
3415 ir_node *add_r = get_Add_right(a);
3416 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3417 /* ~(x + -1) = -x */
3418 ir_node *op = get_Add_left(a);
3419 ir_node *blk = get_nodes_block(n);
3420 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, get_irn_mode(n));
3421 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3429 * Transform a Minus.
3433 * -(a >>u (size-1)) = a >>s (size-1)
3434 * -(a >>s (size-1)) = a >>u (size-1)
3435 * -(a * const) -> a * -const
3437 static ir_node *transform_node_Minus(ir_node *n)
3439 ir_node *c, *oldn = n;
3440 ir_node *a = get_Minus_op(n);
3443 HANDLE_UNOP_PHI(tarval_neg,a,c);
3445 mode = get_irn_mode(a);
3446 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3447 /* the following rules are only to twos-complement */
3450 ir_node *op = get_Not_op(a);
3451 ir_tarval *tv = get_mode_one(mode);
3452 ir_node *blk = get_nodes_block(n);
3453 ir_graph *irg = get_irn_irg(blk);
3454 ir_node *c = new_r_Const(irg, tv);
3455 n = new_rd_Add(get_irn_dbg_info(n), blk, op, c, mode);
3456 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3460 ir_node *c = get_Shr_right(a);
3463 ir_tarval *tv = get_Const_tarval(c);
3465 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3466 /* -(a >>u (size-1)) = a >>s (size-1) */
3467 ir_node *v = get_Shr_left(a);
3469 n = new_rd_Shrs(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3470 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3476 ir_node *c = get_Shrs_right(a);
3479 ir_tarval *tv = get_Const_tarval(c);
3481 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3482 /* -(a >>s (size-1)) = a >>u (size-1) */
3483 ir_node *v = get_Shrs_left(a);
3485 n = new_rd_Shr(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3486 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3493 /* - (a-b) = b - a */
3494 ir_node *la = get_Sub_left(a);
3495 ir_node *ra = get_Sub_right(a);
3496 ir_node *blk = get_nodes_block(n);
3498 n = new_rd_Sub(get_irn_dbg_info(n), blk, ra, la, mode);
3499 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3503 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3504 ir_node *mul_l = get_Mul_left(a);
3505 ir_node *mul_r = get_Mul_right(a);
3506 ir_tarval *tv = value_of(mul_r);
3507 if (tv != tarval_bad) {
3508 tv = tarval_neg(tv);
3509 if (tv != tarval_bad) {
3510 ir_graph *irg = get_irn_irg(n);
3511 ir_node *cnst = new_r_Const(irg, tv);
3512 dbg_info *dbg = get_irn_dbg_info(a);
3513 ir_node *block = get_nodes_block(a);
3514 n = new_rd_Mul(dbg, block, mul_l, cnst, mode);
3515 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3522 } /* transform_node_Minus */
3525 * Transform a Proj(Load) with a non-null address.
3527 static ir_node *transform_node_Proj_Load(ir_node *proj)
3529 if (get_opt_ldst_only_null_ptr_exceptions()) {
3530 if (get_irn_mode(proj) == mode_X) {
3531 ir_node *load = get_Proj_pred(proj);
3533 /* get the Load address */
3534 const ir_node *addr = get_Load_ptr(load);
3535 const ir_node *confirm;
3537 if (value_not_null(addr, &confirm)) {
3538 if (confirm == NULL) {
3539 /* this node may float if it did not depend on a Confirm */
3540 set_irn_pinned(load, op_pin_state_floats);
3542 if (get_Proj_proj(proj) == pn_Load_X_except) {
3543 ir_graph *irg = get_irn_irg(proj);
3544 DBG_OPT_EXC_REM(proj);
3545 return get_irg_bad(irg);
3547 ir_node *blk = get_nodes_block(load);
3548 return new_r_Jmp(blk);
3554 } /* transform_node_Proj_Load */
3557 * Transform a Proj(Store) with a non-null address.
3559 static ir_node *transform_node_Proj_Store(ir_node *proj)
3561 if (get_opt_ldst_only_null_ptr_exceptions()) {
3562 if (get_irn_mode(proj) == mode_X) {
3563 ir_node *store = get_Proj_pred(proj);
3565 /* get the load/store address */
3566 const ir_node *addr = get_Store_ptr(store);
3567 const ir_node *confirm;
3569 if (value_not_null(addr, &confirm)) {
3570 if (confirm == NULL) {
3571 /* this node may float if it did not depend on a Confirm */
3572 set_irn_pinned(store, op_pin_state_floats);
3574 if (get_Proj_proj(proj) == pn_Store_X_except) {
3575 ir_graph *irg = get_irn_irg(proj);
3576 DBG_OPT_EXC_REM(proj);
3577 return get_irg_bad(irg);
3579 ir_node *blk = get_nodes_block(store);
3580 return new_r_Jmp(blk);
3586 } /* transform_node_Proj_Store */
3589 * Transform a Proj(Div) with a non-zero value.
3590 * Removes the exceptions and routes the memory to the NoMem node.
3592 static ir_node *transform_node_Proj_Div(ir_node *proj)
3594 ir_node *div = get_Proj_pred(proj);
3595 ir_node *b = get_Div_right(div);
3596 ir_node *res, *new_mem;
3597 const ir_node *confirm;
3600 if (value_not_zero(b, &confirm)) {
3601 /* div(x, y) && y != 0 */
3602 if (confirm == NULL) {
3603 /* we are sure we have a Const != 0 */
3604 new_mem = get_Div_mem(div);
3605 new_mem = skip_Pin(new_mem);
3606 set_Div_mem(div, new_mem);
3607 set_irn_pinned(div, op_pin_state_floats);
3610 proj_nr = get_Proj_proj(proj);
3612 case pn_Div_X_regular:
3613 return new_r_Jmp(get_nodes_block(div));
3615 case pn_Div_X_except: {
3616 ir_graph *irg = get_irn_irg(proj);
3617 /* we found an exception handler, remove it */
3618 DBG_OPT_EXC_REM(proj);
3619 return get_irg_bad(irg);
3623 ir_graph *irg = get_irn_irg(proj);
3624 res = get_Div_mem(div);
3625 new_mem = get_irg_no_mem(irg);
3628 /* This node can only float up to the Confirm block */
3629 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3631 set_irn_pinned(div, op_pin_state_floats);
3632 /* this is a Div without exception, we can remove the memory edge */
3633 set_Div_mem(div, new_mem);
3639 } /* transform_node_Proj_Div */
3642 * Transform a Proj(Mod) with a non-zero value.
3643 * Removes the exceptions and routes the memory to the NoMem node.
3645 static ir_node *transform_node_Proj_Mod(ir_node *proj)
3647 ir_node *mod = get_Proj_pred(proj);
3648 ir_node *b = get_Mod_right(mod);
3649 ir_node *res, *new_mem;
3650 const ir_node *confirm;
3653 if (value_not_zero(b, &confirm)) {
3654 /* mod(x, y) && y != 0 */
3655 proj_nr = get_Proj_proj(proj);
3657 if (confirm == NULL) {
3658 /* we are sure we have a Const != 0 */
3659 new_mem = get_Mod_mem(mod);
3660 new_mem = skip_Pin(new_mem);
3661 set_Mod_mem(mod, new_mem);
3662 set_irn_pinned(mod, op_pin_state_floats);
3667 case pn_Mod_X_regular:
3668 return new_r_Jmp(get_irn_n(mod, -1));
3670 case pn_Mod_X_except: {
3671 ir_graph *irg = get_irn_irg(proj);
3672 /* we found an exception handler, remove it */
3673 DBG_OPT_EXC_REM(proj);
3674 return get_irg_bad(irg);
3678 ir_graph *irg = get_irn_irg(proj);
3679 res = get_Mod_mem(mod);
3680 new_mem = get_irg_no_mem(irg);
3683 /* This node can only float up to the Confirm block */
3684 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3686 /* this is a Mod without exception, we can remove the memory edge */
3687 set_Mod_mem(mod, new_mem);
3691 if (get_Mod_left(mod) == b) {
3692 /* a % a = 0 if a != 0 */
3693 ir_graph *irg = get_irn_irg(proj);
3694 ir_mode *mode = get_irn_mode(proj);
3695 ir_node *res = new_r_Const(irg, get_mode_null(mode));
3697 DBG_OPT_CSTEVAL(mod, res);
3703 } /* transform_node_Proj_Mod */
3706 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3708 static ir_node *transform_node_Proj_Cond(ir_node *proj)
3710 ir_node *n = get_Proj_pred(proj);
3711 ir_node *b = get_Cond_selector(n);
3713 if (!get_opt_unreachable_code())
3716 if (mode_is_int(get_irn_mode(b))) {
3717 ir_tarval *tb = value_of(b);
3719 if (tb != tarval_bad) {
3720 /* we have a constant switch */
3721 long num = get_Proj_proj(proj);
3723 if (num != get_Cond_default_proj(n)) { /* we cannot optimize default Proj's yet */
3724 if (get_tarval_long(tb) == num) {
3725 /* Do NOT create a jump here, or we will have 2 control flow ops
3726 * in a block. This case is optimized away in optimize_cf(). */
3729 ir_graph *irg = get_irn_irg(proj);
3730 /* this case will NEVER be taken, kill it */
3731 return get_irg_bad(irg);
3735 long num = get_Proj_proj(proj);
3736 vrp_attr *b_vrp = vrp_get_info(b);
3737 if (num != get_Cond_default_proj(n) && b_vrp) {
3738 /* Try handling with vrp data. We only remove dead parts. */
3739 ir_tarval *tp = new_tarval_from_long(num, get_irn_mode(b));
3741 if (b_vrp->range_type == VRP_RANGE) {
3742 ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
3743 ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
3745 if ((cmp_result & ir_relation_greater) == cmp_result && (cmp_result2
3746 & ir_relation_less) == cmp_result2) {
3747 ir_graph *irg = get_irn_irg(proj);
3748 return get_irg_bad(irg);
3750 } else if (b_vrp->range_type == VRP_ANTIRANGE) {
3751 ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
3752 ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
3754 if ((cmp_result & ir_relation_less_equal) == cmp_result && (cmp_result2
3755 & ir_relation_greater_equal) == cmp_result2) {
3756 ir_graph *irg = get_irn_irg(proj);
3757 return get_irg_bad(irg);
3762 tarval_and( b_vrp->bits_set, tp),
3764 ) == ir_relation_equal)) {
3765 ir_graph *irg = get_irn_irg(proj);
3766 return get_irg_bad(irg);
3772 tarval_not(b_vrp->bits_not_set)),
3773 tarval_not(b_vrp->bits_not_set))
3774 == ir_relation_equal)) {
3775 ir_graph *irg = get_irn_irg(proj);
3776 return get_irg_bad(irg);
3787 * return true if the operation returns a value with exactly 1 bit set
3789 static bool is_single_bit(const ir_node *node)
3791 /* a first implementation, could be extended with vrp and others... */
3793 ir_node *shl_l = get_Shl_left(node);
3794 ir_mode *mode = get_irn_mode(node);
3795 int modulo = get_mode_modulo_shift(mode);
3796 /* this works if we shift a 1 and we have modulo shift */
3797 if (is_Const(shl_l) && is_Const_one(shl_l)
3798 && 0 < modulo && modulo <= (int)get_mode_size_bits(mode)) {
3801 } else if (is_Const(node)) {
3802 ir_tarval *tv = get_Const_tarval(node);
3803 return tarval_is_single_bit(tv);
3809 * Normalizes and optimizes Cmp nodes.
3811 static ir_node *transform_node_Cmp(ir_node *n)
3813 ir_node *left = get_Cmp_left(n);
3814 ir_node *right = get_Cmp_right(n);
3815 ir_mode *mode = get_irn_mode(left);
3816 ir_tarval *tv = NULL;
3817 bool changed = false;
3818 bool changedc = false;
3819 ir_relation relation = get_Cmp_relation(n);
3820 ir_relation possible = ir_get_possible_cmp_relations(left, right);
3822 /* mask out impossible relations */
3823 ir_relation new_relation = relation & possible;
3824 if (new_relation != relation) {
3825 relation = new_relation;
3829 /* Remove unnecessary conversions */
3830 /* TODO handle conv+constant */
3831 if (is_Conv(left) && is_Conv(right)) {
3832 ir_node *op_left = get_Conv_op(left);
3833 ir_node *op_right = get_Conv_op(right);
3834 ir_mode *mode_left = get_irn_mode(op_left);
3835 ir_mode *mode_right = get_irn_mode(op_right);
3837 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3838 && mode_left != mode_b && mode_right != mode_b) {
3839 ir_node *block = get_nodes_block(n);
3841 if (mode_left == mode_right) {
3845 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
3846 } else if (smaller_mode(mode_left, mode_right)) {
3847 left = new_r_Conv(block, op_left, mode_right);
3850 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3851 } else if (smaller_mode(mode_right, mode_left)) {
3853 right = new_r_Conv(block, op_right, mode_left);
3855 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3861 * Optimize -a CMP -b into b CMP a.
3862 * This works only for modes where unary Minus cannot Overflow.
3863 * Note that two-complement integers can Overflow so it will NOT work.
3865 if (!mode_overflow_on_unary_Minus(mode) &&
3866 is_Minus(left) && is_Minus(right)) {
3867 left = get_Minus_op(left);
3868 right = get_Minus_op(right);
3869 relation = get_inversed_relation(relation);
3871 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3874 /* remove operation on both sides if possible */
3875 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
3877 * The following operations are NOT safe for floating point operations, for instance
3878 * 1.0 + inf == 2.0 + inf, =/=> x == y
3880 if (mode_is_int(mode)) {
3881 unsigned lop = get_irn_opcode(left);
3883 if (lop == get_irn_opcode(right)) {
3884 ir_node *ll, *lr, *rl, *rr;
3886 /* same operation on both sides, try to remove */
3890 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
3891 left = get_unop_op(left);
3892 right = get_unop_op(right);
3894 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3897 ll = get_Add_left(left);
3898 lr = get_Add_right(left);
3899 rl = get_Add_left(right);
3900 rr = get_Add_right(right);
3903 /* X + a CMP X + b ==> a CMP b */
3907 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3908 } else if (ll == rr) {
3909 /* X + a CMP b + X ==> a CMP b */
3913 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3914 } else if (lr == rl) {
3915 /* a + X CMP X + b ==> a CMP b */
3919 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3920 } else if (lr == rr) {
3921 /* a + X CMP b + X ==> a CMP b */
3925 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3929 ll = get_Sub_left(left);
3930 lr = get_Sub_right(left);
3931 rl = get_Sub_left(right);
3932 rr = get_Sub_right(right);
3935 /* X - a CMP X - b ==> a CMP b */
3939 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3940 } else if (lr == rr) {
3941 /* a - X CMP b - X ==> a CMP b */
3945 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3949 if (get_Rotl_right(left) == get_Rotl_right(right)) {
3950 /* a ROTL X CMP b ROTL X ==> a CMP b */
3951 left = get_Rotl_left(left);
3952 right = get_Rotl_left(right);
3954 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3962 /* X+A == A, A+X == A, A-X == A -> X == 0 */
3963 if (is_Add(left) || is_Sub(left)) {
3964 ir_node *ll = get_binop_left(left);
3965 ir_node *lr = get_binop_right(left);
3967 if (lr == right && is_Add(left)) {
3973 ir_graph *irg = get_irn_irg(n);
3975 right = create_zero_const(irg, mode);
3977 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3980 if (is_Add(right) || is_Sub(right)) {
3981 ir_node *rl = get_binop_left(right);
3982 ir_node *rr = get_binop_right(right);
3984 if (rr == left && is_Add(right)) {
3990 ir_graph *irg = get_irn_irg(n);
3992 right = create_zero_const(irg, mode);
3994 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3998 if (is_And(left) && is_Const(right)) {
3999 ir_node *ll = get_binop_left(left);
4000 ir_node *lr = get_binop_right(left);
4001 if (is_Shr(ll) && is_Const(lr)) {
4002 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
4003 ir_node *block = get_nodes_block(n);
4004 ir_mode *mode = get_irn_mode(left);
4006 ir_node *llr = get_Shr_right(ll);
4007 if (is_Const(llr)) {
4008 dbg_info *dbg = get_irn_dbg_info(left);
4009 ir_graph *irg = get_irn_irg(left);
4011 ir_tarval *c1 = get_Const_tarval(llr);
4012 ir_tarval *c2 = get_Const_tarval(lr);
4013 ir_tarval *c3 = get_Const_tarval(right);
4014 ir_tarval *mask = tarval_shl(c2, c1);
4015 ir_tarval *value = tarval_shl(c3, c1);
4017 left = new_rd_And(dbg, block, get_Shr_left(ll), new_r_Const(irg, mask), mode);
4018 right = new_r_Const(irg, value);
4023 /* Cmp(Eor(x, y), 0) <=> Cmp(x, y) at least for the ==0,!=0
4025 if (is_Const(right) && is_Const_null(right) && is_Eor(left)) {
4026 right = get_Eor_right(left);
4027 left = get_Eor_left(left);
4030 } /* mode_is_int(...) */
4033 /* Cmp(And(1bit, val), 1bit) "bit-testing" can be replaced
4034 * by the simpler Cmp(And(1bit), val), 0) negated pnc */
4035 if (mode_is_int(mode) && is_And(left)
4036 && (relation == ir_relation_equal
4037 || (mode_is_signed(mode) && relation == ir_relation_less_greater)
4038 || (!mode_is_signed(mode) && (relation & ir_relation_less_equal) == ir_relation_less))) {
4039 ir_node *and0 = get_And_left(left);
4040 ir_node *and1 = get_And_right(left);
4041 if (and1 == right) {
4042 ir_node *tmp = and0;
4046 if (and0 == right && is_single_bit(and0)) {
4047 ir_graph *irg = get_irn_irg(n);
4049 relation == ir_relation_equal ? ir_relation_less_greater : ir_relation_equal;
4050 right = create_zero_const(irg, mode);
4055 /* replace mode_b compares with ands/ors */
4056 if (mode == mode_b) {
4057 ir_node *block = get_nodes_block(n);
4061 case ir_relation_less_equal:
4062 bres = new_r_Or(block, new_r_Not(block, left, mode_b), right, mode_b);
4064 case ir_relation_less:
4065 bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b);
4067 case ir_relation_greater_equal:
4068 bres = new_r_Or(block, left, new_r_Not(block, right, mode_b), mode_b);
4070 case ir_relation_greater:
4071 bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b);
4073 case ir_relation_less_greater:
4074 bres = new_r_Eor(block, left, right, mode_b);
4076 case ir_relation_equal:
4077 bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b);
4080 #ifdef DEBUG_libfirm
4081 ir_fprintf(stderr, "Optimisation warning, unexpected mode_b Cmp %+F\n", n);
4086 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4092 * First step: normalize the compare op
4093 * by placing the constant on the right side
4094 * or moving the lower address node to the left.
4096 if (!operands_are_normalized(left, right)) {
4101 relation = get_inversed_relation(relation);
4106 * Second step: Try to reduce the magnitude
4107 * of a constant. This may help to generate better code
4108 * later and may help to normalize more compares.
4109 * Of course this is only possible for integer values.
4111 tv = value_of(right);
4112 if (tv != tarval_bad) {
4113 ir_mode *mode = get_irn_mode(right);
4115 /* TODO extend to arbitrary constants */
4116 if (is_Conv(left) && tarval_is_null(tv)) {
4117 ir_node *op = get_Conv_op(left);
4118 ir_mode *op_mode = get_irn_mode(op);
4121 * UpConv(x) REL 0 ==> x REL 0
4122 * Don't do this for float values as it's unclear whether it is a
4123 * win. (on the other side it makes detection/creation of fabs hard)
4125 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4126 ((relation == ir_relation_equal || relation == ir_relation_less_greater) ||
4127 mode_is_signed(mode) || !mode_is_signed(op_mode)) &&
4128 !mode_is_float(mode)) {
4129 tv = get_mode_null(op_mode);
4133 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4137 if (tv != tarval_bad) {
4138 /* the following optimization is possible on modes without Overflow
4139 * on Unary Minus or on == and !=:
4140 * -a CMP c ==> a swap(CMP) -c
4142 * Beware: for two-complement Overflow may occur, so only == and != can
4143 * be optimized, see this:
4144 * -MININT < 0 =/=> MININT > 0 !!!
4146 if (is_Minus(left) &&
4147 (!mode_overflow_on_unary_Minus(mode) ||
4148 (mode_is_int(mode) && (relation == ir_relation_equal || relation == ir_relation_less_greater)))) {
4149 tv = tarval_neg(tv);
4151 if (tv != tarval_bad) {
4152 left = get_Minus_op(left);
4153 relation = get_inversed_relation(relation);
4155 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4157 } else if (is_Not(left) && (relation == ir_relation_equal || relation == ir_relation_less_greater)) {
4158 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4159 tv = tarval_not(tv);
4161 if (tv != tarval_bad) {
4162 left = get_Not_op(left);
4164 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4168 /* for integer modes, we have more */
4169 if (mode_is_int(mode)) {
4170 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4171 if ((relation == ir_relation_less || relation == ir_relation_greater_equal) &&
4172 tarval_cmp(tv, get_mode_null(mode)) == ir_relation_greater) {
4173 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4175 if (tv != tarval_bad) {
4176 relation ^= ir_relation_equal;
4178 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4181 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4182 else if ((relation == ir_relation_greater || relation == ir_relation_less_equal) &&
4183 tarval_cmp(tv, get_mode_null(mode)) == ir_relation_less) {
4184 tv = tarval_add(tv, get_mode_one(mode));
4186 if (tv != tarval_bad) {
4187 relation ^= ir_relation_equal;
4189 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4193 /* the following reassociations work only for == and != */
4194 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4196 #if 0 /* Might be not that good in general */
4197 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4198 if (tarval_is_null(tv) && is_Sub(left)) {
4199 right = get_Sub_right(left);
4200 left = get_Sub_left(left);
4202 tv = value_of(right);
4204 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4208 if (tv != tarval_bad) {
4209 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4211 ir_node *c1 = get_Sub_right(left);
4212 ir_tarval *tv2 = value_of(c1);
4214 if (tv2 != tarval_bad) {
4215 tv2 = tarval_add(tv, value_of(c1));
4217 if (tv2 != tarval_bad) {
4218 left = get_Sub_left(left);
4221 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4225 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4226 else if (is_Add(left)) {
4227 ir_node *a_l = get_Add_left(left);
4228 ir_node *a_r = get_Add_right(left);
4232 if (is_Const(a_l)) {
4234 tv2 = value_of(a_l);
4237 tv2 = value_of(a_r);
4240 if (tv2 != tarval_bad) {
4241 tv2 = tarval_sub(tv, tv2, NULL);
4243 if (tv2 != tarval_bad) {
4247 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4251 /* -a == c ==> a == -c, -a != c ==> a != -c */
4252 else if (is_Minus(left)) {
4253 ir_tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4255 if (tv2 != tarval_bad) {
4256 left = get_Minus_op(left);
4259 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4266 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4267 switch (get_irn_opcode(left)) {
4271 c1 = get_And_right(left);
4274 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4275 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4277 ir_tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4279 /* TODO: move to constant evaluation */
4280 ir_graph *irg = get_irn_irg(n);
4281 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4282 c1 = new_r_Const(irg, tv);
4283 DBG_OPT_CSTEVAL(n, c1);
4287 if (tarval_is_single_bit(tv)) {
4289 * optimization for AND:
4291 * And(x, C) == C ==> And(x, C) != 0
4292 * And(x, C) != C ==> And(X, C) == 0
4294 * if C is a single Bit constant.
4297 /* check for Constant's match. We have check hare the tarvals,
4298 because our const might be changed */
4299 if (get_Const_tarval(c1) == tv) {
4300 /* fine: do the transformation */
4301 tv = get_mode_null(get_tarval_mode(tv));
4302 relation ^= ir_relation_less_equal_greater;
4304 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4310 c1 = get_Or_right(left);
4311 if (is_Const(c1) && tarval_is_null(tv)) {
4313 * Or(x, C) == 0 && C != 0 ==> FALSE
4314 * Or(x, C) != 0 && C != 0 ==> TRUE
4316 if (! tarval_is_null(get_Const_tarval(c1))) {
4317 /* TODO: move to constant evaluation */
4318 ir_graph *irg = get_irn_irg(n);
4319 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4320 c1 = new_r_Const(irg, tv);
4321 DBG_OPT_CSTEVAL(n, c1);
4328 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4330 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4333 c1 = get_Shl_right(left);
4335 ir_graph *irg = get_irn_irg(c1);
4336 ir_tarval *tv1 = get_Const_tarval(c1);
4337 ir_mode *mode = get_irn_mode(left);
4338 ir_tarval *minus1 = get_mode_all_one(mode);
4339 ir_tarval *amask = tarval_shr(minus1, tv1);
4340 ir_tarval *cmask = tarval_shl(minus1, tv1);
4343 if (tarval_and(tv, cmask) != tv) {
4344 /* condition not met */
4345 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4346 c1 = new_r_Const(irg, tv);
4347 DBG_OPT_CSTEVAL(n, c1);
4350 sl = get_Shl_left(left);
4351 blk = get_nodes_block(n);
4352 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4353 tv = tarval_shr(tv, tv1);
4355 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4360 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4362 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4365 c1 = get_Shr_right(left);
4367 ir_graph *irg = get_irn_irg(c1);
4368 ir_tarval *tv1 = get_Const_tarval(c1);
4369 ir_mode *mode = get_irn_mode(left);
4370 ir_tarval *minus1 = get_mode_all_one(mode);
4371 ir_tarval *amask = tarval_shl(minus1, tv1);
4372 ir_tarval *cmask = tarval_shr(minus1, tv1);
4375 if (tarval_and(tv, cmask) != tv) {
4376 /* condition not met */
4377 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4378 c1 = new_r_Const(irg, tv);
4379 DBG_OPT_CSTEVAL(n, c1);
4382 sl = get_Shr_left(left);
4383 blk = get_nodes_block(n);
4384 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4385 tv = tarval_shl(tv, tv1);
4387 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4392 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4394 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4397 c1 = get_Shrs_right(left);
4399 ir_graph *irg = get_irn_irg(c1);
4400 ir_tarval *tv1 = get_Const_tarval(c1);
4401 ir_mode *mode = get_irn_mode(left);
4402 ir_tarval *minus1 = get_mode_all_one(mode);
4403 ir_tarval *amask = tarval_shl(minus1, tv1);
4404 ir_tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4407 cond = tarval_sub(cond, tv1, NULL);
4408 cond = tarval_shrs(tv, cond);
4410 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4411 /* condition not met */
4412 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4413 c1 = new_r_Const(irg, tv);
4414 DBG_OPT_CSTEVAL(n, c1);
4417 sl = get_Shrs_left(left);
4418 blk = get_nodes_block(n);
4419 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4420 tv = tarval_shl(tv, tv1);
4422 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4427 } /* tarval != bad */
4430 if (changedc) { /* need a new Const */
4431 ir_graph *irg = get_irn_irg(n);
4432 right = new_r_Const(irg, tv);
4436 if ((relation == ir_relation_equal || relation == ir_relation_less_greater) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4437 ir_node *op = get_Proj_pred(left);
4439 if (is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) {
4440 ir_node *c = get_binop_right(op);
4443 ir_tarval *tv = get_Const_tarval(c);
4445 if (tarval_is_single_bit(tv)) {
4446 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4447 ir_node *v = get_binop_left(op);
4448 ir_node *blk = get_irn_n(op, -1);
4449 ir_graph *irg = get_irn_irg(op);
4450 ir_mode *mode = get_irn_mode(v);
4452 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4453 left = new_rd_And(get_irn_dbg_info(op), blk, v, new_r_Const(irg, tv), mode);
4455 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4462 dbg_info *dbgi = get_irn_dbg_info(n);
4463 ir_node *block = get_nodes_block(n);
4465 /* create a new compare */
4466 n = new_rd_Cmp(dbgi, block, left, right, relation);
4473 * Optimize CopyB(mem, x, x) into a Nop.
4475 static ir_node *transform_node_Proj_CopyB(ir_node *proj)
4477 ir_node *copyb = get_Proj_pred(proj);
4478 ir_node *a = get_CopyB_dst(copyb);
4479 ir_node *b = get_CopyB_src(copyb);
4482 switch (get_Proj_proj(proj)) {
4483 case pn_CopyB_X_regular:
4484 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4485 DBG_OPT_EXC_REM(proj);
4486 proj = new_r_Jmp(get_nodes_block(copyb));
4488 case pn_CopyB_X_except:
4489 DBG_OPT_EXC_REM(proj);
4490 proj = get_irg_bad(get_irn_irg(proj));
4497 } /* transform_node_Proj_CopyB */
4500 * Optimize Bounds(idx, idx, upper) into idx.
4502 static ir_node *transform_node_Proj_Bound(ir_node *proj)
4504 ir_node *oldn = proj;
4505 ir_node *bound = get_Proj_pred(proj);
4506 ir_node *idx = get_Bound_index(bound);
4507 ir_node *pred = skip_Proj(idx);
4510 if (idx == get_Bound_lower(bound))
4512 else if (is_Bound(pred)) {
4514 * idx was Bounds checked previously, it is still valid if
4515 * lower <= pred_lower && pred_upper <= upper.
4517 ir_node *lower = get_Bound_lower(bound);
4518 ir_node *upper = get_Bound_upper(bound);
4519 if (get_Bound_lower(pred) == lower &&
4520 get_Bound_upper(pred) == upper) {
4522 * One could expect that we simply return the previous
4523 * Bound here. However, this would be wrong, as we could
4524 * add an exception Proj to a new location then.
4525 * So, we must turn in into a tuple.
4531 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4532 switch (get_Proj_proj(proj)) {
4534 DBG_OPT_EXC_REM(proj);
4535 proj = get_Bound_mem(bound);
4537 case pn_Bound_X_except:
4538 DBG_OPT_EXC_REM(proj);
4539 proj = get_irg_bad(get_irn_irg(proj));
4543 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4545 case pn_Bound_X_regular:
4546 DBG_OPT_EXC_REM(proj);
4547 proj = new_r_Jmp(get_nodes_block(bound));
4554 } /* transform_node_Proj_Bound */
4557 * Does all optimizations on nodes that must be done on its Projs
4558 * because of creating new nodes.
4560 static ir_node *transform_node_Proj(ir_node *proj)
4562 ir_node *n = get_Proj_pred(proj);
4564 if (n->op->ops.transform_node_Proj)
4565 return n->op->ops.transform_node_Proj(proj);
4567 } /* transform_node_Proj */
4570 * Move Confirms down through Phi nodes.
4572 static ir_node *transform_node_Phi(ir_node *phi)
4575 ir_mode *mode = get_irn_mode(phi);
4577 if (mode_is_reference(mode)) {
4578 n = get_irn_arity(phi);
4580 /* Beware of Phi0 */
4582 ir_node *pred = get_irn_n(phi, 0);
4583 ir_node *bound, *new_phi, *block, **in;
4584 ir_relation relation;
4586 if (! is_Confirm(pred))
4589 bound = get_Confirm_bound(pred);
4590 relation = get_Confirm_relation(pred);
4592 NEW_ARR_A(ir_node *, in, n);
4593 in[0] = get_Confirm_value(pred);
4595 for (i = 1; i < n; ++i) {
4596 pred = get_irn_n(phi, i);
4598 if (! is_Confirm(pred) ||
4599 get_Confirm_bound(pred) != bound ||
4600 get_Confirm_relation(pred) != relation)
4602 in[i] = get_Confirm_value(pred);
4604 /* move the Confirm nodes "behind" the Phi */
4605 block = get_irn_n(phi, -1);
4606 new_phi = new_r_Phi(block, n, in, get_irn_mode(phi));
4607 return new_r_Confirm(block, new_phi, bound, relation);
4614 * Returns the operands of a commutative bin-op, if one operand is
4615 * a const, it is returned as the second one.
4617 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c)
4619 ir_node *op_a = get_binop_left(binop);
4620 ir_node *op_b = get_binop_right(binop);
4622 assert(is_op_commutative(get_irn_op(binop)));
4624 if (is_Const(op_a)) {
4631 } /* get_comm_Binop_Ops */
4634 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4635 * Such pattern may arise in bitfield stores.
4637 * value c4 value c4 & c2
4638 * AND c3 AND c1 | c3
4645 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4648 static ir_node *transform_node_Or_bf_store(ir_node *irn_or)
4650 ir_node *irn_and, *c1;
4652 ir_node *and_l, *c3;
4653 ir_node *value, *c4;
4654 ir_node *new_and, *new_const, *block;
4655 ir_mode *mode = get_irn_mode(irn_or);
4657 ir_tarval *tv1, *tv2, *tv3, *tv4, *tv;
4661 get_comm_Binop_Ops(irn_or, &irn_and, &c1);
4662 if (!is_Const(c1) || !is_And(irn_and))
4665 get_comm_Binop_Ops(irn_and, &or_l, &c2);
4669 tv1 = get_Const_tarval(c1);
4670 tv2 = get_Const_tarval(c2);
4672 tv = tarval_or(tv1, tv2);
4673 if (tarval_is_all_one(tv)) {
4674 /* the AND does NOT clear a bit with isn't set by the OR */
4675 set_Or_left(irn_or, or_l);
4676 set_Or_right(irn_or, c1);
4678 /* check for more */
4685 get_comm_Binop_Ops(or_l, &and_l, &c3);
4686 if (!is_Const(c3) || !is_And(and_l))
4689 get_comm_Binop_Ops(and_l, &value, &c4);
4693 /* ok, found the pattern, check for conditions */
4694 assert(mode == get_irn_mode(irn_and));
4695 assert(mode == get_irn_mode(or_l));
4696 assert(mode == get_irn_mode(and_l));
4698 tv3 = get_Const_tarval(c3);
4699 tv4 = get_Const_tarval(c4);
4701 tv = tarval_or(tv4, tv2);
4702 if (!tarval_is_all_one(tv)) {
4703 /* have at least one 0 at the same bit position */
4707 if (tv3 != tarval_andnot(tv3, tv4)) {
4708 /* bit in the or_mask is outside the and_mask */
4712 if (tv1 != tarval_andnot(tv1, tv2)) {
4713 /* bit in the or_mask is outside the and_mask */
4717 /* ok, all conditions met */
4718 block = get_irn_n(irn_or, -1);
4719 irg = get_irn_irg(block);
4721 new_and = new_r_And(block, value, new_r_Const(irg, tarval_and(tv4, tv2)), mode);
4723 new_const = new_r_Const(irg, tarval_or(tv3, tv1));
4725 set_Or_left(irn_or, new_and);
4726 set_Or_right(irn_or, new_const);
4728 /* check for more */
4730 } /* transform_node_Or_bf_store */
4733 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4735 static ir_node *transform_node_Or_Rotl(ir_node *irn_or)
4737 ir_mode *mode = get_irn_mode(irn_or);
4738 ir_node *shl, *shr, *block;
4739 ir_node *irn, *x, *c1, *c2, *n;
4740 ir_tarval *tv1, *tv2;
4742 /* some backends can't handle rotl */
4743 if (!be_get_backend_param()->support_rotl)
4746 if (! mode_is_int(mode))
4749 shl = get_binop_left(irn_or);
4750 shr = get_binop_right(irn_or);
4759 } else if (!is_Shl(shl)) {
4761 } else if (!is_Shr(shr)) {
4764 x = get_Shl_left(shl);
4765 if (x != get_Shr_left(shr))
4768 c1 = get_Shl_right(shl);
4769 c2 = get_Shr_right(shr);
4770 if (is_Const(c1) && is_Const(c2)) {
4771 tv1 = get_Const_tarval(c1);
4772 if (! tarval_is_long(tv1))
4775 tv2 = get_Const_tarval(c2);
4776 if (! tarval_is_long(tv2))
4779 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4780 != (int) get_mode_size_bits(mode))
4783 /* yet, condition met */
4784 block = get_nodes_block(irn_or);
4786 n = new_r_Rotl(block, x, c1, mode);
4788 DBG_OPT_ALGSIM1(irn_or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4792 /* Note: the obvious rot formulation (a << x) | (a >> (32-x)) gets
4793 * transformed to (a << x) | (a >> -x) by transform_node_shift_modulo() */
4794 if (!ir_is_negated_value(c1, c2)) {
4798 /* yet, condition met */
4799 block = get_nodes_block(irn_or);
4800 n = new_r_Rotl(block, x, c1, mode);
4801 DBG_OPT_ALGSIM0(irn_or, n, FS_OPT_OR_SHFT_TO_ROTL);
4803 } /* transform_node_Or_Rotl */
4805 static bool is_cmp_unequal_zero(const ir_node *node)
4807 ir_relation relation = get_Cmp_relation(node);
4808 ir_node *left = get_Cmp_left(node);
4809 ir_node *right = get_Cmp_right(node);
4810 ir_mode *mode = get_irn_mode(left);
4812 if (!is_Const(right) || !is_Const_null(right))
4814 if (mode_is_signed(mode)) {
4815 return relation == ir_relation_less_greater;
4817 return relation == ir_relation_greater;
4824 static ir_node *transform_node_Or(ir_node *n)
4826 ir_node *c, *oldn = n;
4827 ir_node *a = get_Or_left(n);
4828 ir_node *b = get_Or_right(n);
4831 if (is_Not(a) && is_Not(b)) {
4832 /* ~a | ~b = ~(a&b) */
4833 ir_node *block = get_nodes_block(n);
4835 mode = get_irn_mode(n);
4838 n = new_rd_And(get_irn_dbg_info(n), block, a, b, mode);
4839 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
4840 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4844 /* we can combine the relations of two compares with the same operands */
4845 if (is_Cmp(a) && is_Cmp(b)) {
4846 ir_node *a_left = get_Cmp_left(a);
4847 ir_node *a_right = get_Cmp_left(a);
4848 ir_node *b_left = get_Cmp_left(b);
4849 ir_node *b_right = get_Cmp_right(b);
4850 if (a_left == b_left && b_left == b_right) {
4851 dbg_info *dbgi = get_irn_dbg_info(n);
4852 ir_node *block = get_nodes_block(n);
4853 ir_relation a_relation = get_Cmp_relation(a);
4854 ir_relation b_relation = get_Cmp_relation(b);
4855 ir_relation new_relation = a_relation | b_relation;
4856 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
4858 /* Cmp(a!=0) or Cmp(b!=0) => Cmp(a|b != 0) */
4859 if (is_cmp_unequal_zero(a) && is_cmp_unequal_zero(b)
4860 && !mode_is_float(get_irn_mode(a_left))
4861 && !mode_is_float(get_irn_mode(b_left))) {
4862 ir_graph *irg = get_irn_irg(n);
4863 dbg_info *dbgi = get_irn_dbg_info(n);
4864 ir_node *block = get_nodes_block(n);
4865 ir_mode *mode = get_irn_mode(a_left);
4866 ir_node *n_b_left = get_irn_mode(b_left) != mode ?
4867 new_rd_Conv(dbgi, block, b_left, mode) : b_left;
4868 ir_node *or = new_rd_Or(dbgi, block, a_left, n_b_left, mode);
4869 ir_node *zero = create_zero_const(irg, mode);
4870 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_less_greater);
4874 mode = get_irn_mode(n);
4875 HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode);
4877 n = transform_node_Or_bf_store(n);
4878 n = transform_node_Or_Rotl(n);
4882 n = transform_bitwise_distributive(n, transform_node_Or);
4885 } /* transform_node_Or */
4889 static ir_node *transform_node(ir_node *n);
4892 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
4894 * Should be moved to reassociation?
4896 static ir_node *transform_node_shift(ir_node *n)
4898 ir_node *left, *right;
4900 ir_tarval *tv1, *tv2, *res;
4901 ir_node *in[2], *irn, *block;
4904 left = get_binop_left(n);
4906 /* different operations */
4907 if (get_irn_op(left) != get_irn_op(n))
4910 right = get_binop_right(n);
4911 tv1 = value_of(right);
4912 if (tv1 == tarval_bad)
4915 tv2 = value_of(get_binop_right(left));
4916 if (tv2 == tarval_bad)
4919 res = tarval_add(tv1, tv2);
4920 mode = get_irn_mode(n);
4921 irg = get_irn_irg(n);
4923 /* beware: a simple replacement works only, if res < modulo shift */
4925 int modulo_shf = get_mode_modulo_shift(mode);
4926 if (modulo_shf > 0) {
4927 ir_tarval *modulo = new_tarval_from_long(modulo_shf,
4928 get_tarval_mode(res));
4930 assert(modulo_shf >= (int) get_mode_size_bits(mode));
4932 /* shifting too much */
4933 if (!(tarval_cmp(res, modulo) & ir_relation_less)) {
4935 ir_node *block = get_nodes_block(n);
4936 dbg_info *dbgi = get_irn_dbg_info(n);
4937 ir_mode *smode = get_irn_mode(right);
4938 ir_node *cnst = new_r_Const_long(irg, smode, get_mode_size_bits(mode) - 1);
4939 return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode);
4942 return new_r_Const(irg, get_mode_null(mode));
4946 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
4949 /* ok, we can replace it */
4950 block = get_nodes_block(n);
4952 in[0] = get_binop_left(left);
4953 in[1] = new_r_Const(irg, res);
4955 irn = new_ir_node(NULL, get_Block_irg(block), block, get_irn_op(n), mode, 2, in);
4957 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
4959 return transform_node(irn);
4960 } /* transform_node_shift */
4963 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
4965 * - and, or, xor instead of &
4966 * - Shl, Shr, Shrs, rotl instead of >>
4967 * (with a special case for Or/Xor + Shrs)
4969 static ir_node *transform_node_bitop_shift(ir_node *n)
4972 ir_node *right = get_binop_right(n);
4973 ir_mode *mode = get_irn_mode(n);
4974 ir_node *bitop_left;
4975 ir_node *bitop_right;
4985 ir_tarval *tv_shift;
4987 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
4989 if (!is_Const(right))
4992 left = get_binop_left(n);
4993 op_left = get_irn_op(left);
4994 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
4997 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
4998 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
4999 /* TODO: test if sign bit is affectes */
5003 bitop_right = get_binop_right(left);
5004 if (!is_Const(bitop_right))
5007 bitop_left = get_binop_left(left);
5009 block = get_nodes_block(n);
5010 dbgi = get_irn_dbg_info(n);
5011 tv1 = get_Const_tarval(bitop_right);
5012 tv2 = get_Const_tarval(right);
5014 assert(get_tarval_mode(tv1) == mode);
5017 new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode);
5018 tv_shift = tarval_shl(tv1, tv2);
5019 } else if (is_Shr(n)) {
5020 new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode);
5021 tv_shift = tarval_shr(tv1, tv2);
5022 } else if (is_Shrs(n)) {
5023 new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode);
5024 tv_shift = tarval_shrs(tv1, tv2);
5027 new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode);
5028 tv_shift = tarval_rotl(tv1, tv2);
5031 assert(get_tarval_mode(tv_shift) == mode);
5032 irg = get_irn_irg(n);
5033 new_const = new_r_Const(irg, tv_shift);
5035 if (op_left == op_And) {
5036 new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode);
5037 } else if (op_left == op_Or) {
5038 new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode);
5040 assert(op_left == op_Eor);
5041 new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode);
5049 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5051 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5052 * (also with x >>s c1 when c1>=c2)
5054 static ir_node *transform_node_shl_shr(ir_node *n)
5057 ir_node *right = get_binop_right(n);
5067 ir_tarval *tv_shift;
5070 ir_relation relation;
5073 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5075 if (!is_Const(right))
5078 left = get_binop_left(n);
5079 mode = get_irn_mode(n);
5080 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5081 ir_node *shr_right = get_binop_right(left);
5083 if (!is_Const(shr_right))
5086 x = get_binop_left(left);
5087 tv_shr = get_Const_tarval(shr_right);
5088 tv_shl = get_Const_tarval(right);
5090 if (is_Shrs(left)) {
5091 /* shrs variant only allowed if c1 >= c2 */
5092 if (! (tarval_cmp(tv_shl, tv_shr) & ir_relation_greater_equal))
5095 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5098 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5100 tv_mask = tarval_shl(tv_mask, tv_shl);
5101 } else if (is_Shr(n) && is_Shl(left)) {
5102 ir_node *shl_right = get_Shl_right(left);
5104 if (!is_Const(shl_right))
5107 x = get_Shl_left(left);
5108 tv_shr = get_Const_tarval(right);
5109 tv_shl = get_Const_tarval(shl_right);
5111 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5112 tv_mask = tarval_shr(tv_mask, tv_shr);
5117 if (get_tarval_mode(tv_shl) != get_tarval_mode(tv_shr)) {
5118 tv_shl = tarval_convert_to(tv_shl, get_tarval_mode(tv_shr));
5121 assert(tv_mask != tarval_bad);
5122 assert(get_tarval_mode(tv_mask) == mode);
5124 block = get_nodes_block(n);
5125 irg = get_irn_irg(block);
5126 dbgi = get_irn_dbg_info(n);
5128 relation = tarval_cmp(tv_shl, tv_shr);
5129 if (relation == ir_relation_less || relation == ir_relation_equal) {
5130 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5131 new_const = new_r_Const(irg, tv_shift);
5133 new_shift = new_rd_Shrs(dbgi, block, x, new_const, mode);
5135 new_shift = new_rd_Shr(dbgi, block, x, new_const, mode);
5138 assert(relation == ir_relation_greater);
5139 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5140 new_const = new_r_Const(irg, tv_shift);
5141 new_shift = new_rd_Shl(dbgi, block, x, new_const, mode);
5144 new_const = new_r_Const(irg, tv_mask);
5145 new_and = new_rd_And(dbgi, block, new_shift, new_const, mode);
5150 static ir_tarval *get_modulo_tv_value(ir_tarval *tv, int modulo_val)
5152 ir_mode *mode = get_tarval_mode(tv);
5153 ir_tarval *modulo_tv = new_tarval_from_long(modulo_val, mode);
5154 return tarval_mod(tv, modulo_tv);
5157 typedef ir_node*(*new_shift_func)(dbg_info *dbgi, ir_node *block,
5158 ir_node *left, ir_node *right, ir_mode *mode);
5161 * Normalisation: if we have a shl/shr with modulo_shift behaviour
5162 * then we can use that to minimize the value of Add(x, const) or
5163 * Sub(Const, x). In particular this often avoids 1 instruction in some
5164 * backends for the Shift(x, Sub(Const, y)) case because it can be replaced
5165 * by Shift(x, Minus(y)) which doesnt't need an explicit Const constructed.
5167 static ir_node *transform_node_shift_modulo(ir_node *n,
5168 new_shift_func new_shift)
5170 ir_mode *mode = get_irn_mode(n);
5171 int modulo = get_mode_modulo_shift(mode);
5172 ir_node *newop = NULL;
5173 ir_mode *mode_right;
5180 if (get_mode_arithmetic(mode) != irma_twos_complement)
5182 if (!is_po2(modulo))
5185 irg = get_irn_irg(n);
5186 block = get_nodes_block(n);
5187 right = get_binop_right(n);
5188 mode_right = get_irn_mode(right);
5189 if (is_Const(right)) {
5190 ir_tarval *tv = get_Const_tarval(right);
5191 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5196 newop = new_r_Const(irg, tv_mod);
5197 } else if (is_Add(right)) {
5198 ir_node *add_right = get_Add_right(right);
5199 if (is_Const(add_right)) {
5200 ir_tarval *tv = get_Const_tarval(add_right);
5201 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5206 newconst = new_r_Const(irg, tv_mod);
5207 newop = new_r_Add(block, get_Add_left(right), newconst,
5210 } else if (is_Sub(right)) {
5211 ir_node *sub_left = get_Sub_left(right);
5212 if (is_Const(sub_left)) {
5213 ir_tarval *tv = get_Const_tarval(sub_left);
5214 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5219 newconst = new_r_Const(irg, tv_mod);
5220 newop = new_r_Sub(block, newconst, get_Sub_right(right),
5227 if (newop != NULL) {
5228 dbg_info *dbgi = get_irn_dbg_info(n);
5229 ir_node *left = get_binop_left(n);
5230 return new_shift(dbgi, block, left, newop, mode);
5238 static ir_node *transform_node_Shr(ir_node *n)
5240 ir_node *c, *oldn = n;
5241 ir_node *left = get_Shr_left(n);
5242 ir_node *right = get_Shr_right(n);
5243 ir_mode *mode = get_irn_mode(n);
5245 HANDLE_BINOP_PHI((eval_func) tarval_shr, left, right, c, mode);
5246 n = transform_node_shift(n);
5249 n = transform_node_shift_modulo(n, new_rd_Shr);
5251 n = transform_node_shl_shr(n);
5253 n = transform_node_bitop_shift(n);
5256 } /* transform_node_Shr */
5261 static ir_node *transform_node_Shrs(ir_node *n)
5263 ir_node *c, *oldn = n;
5264 ir_node *a = get_Shrs_left(n);
5265 ir_node *b = get_Shrs_right(n);
5266 ir_mode *mode = get_irn_mode(n);
5268 HANDLE_BINOP_PHI((eval_func) tarval_shrs, a, b, c, mode);
5269 n = transform_node_shift(n);
5272 n = transform_node_shift_modulo(n, new_rd_Shrs);
5274 n = transform_node_bitop_shift(n);
5277 } /* transform_node_Shrs */
5282 static ir_node *transform_node_Shl(ir_node *n)
5284 ir_node *c, *oldn = n;
5285 ir_node *a = get_Shl_left(n);
5286 ir_node *b = get_Shl_right(n);
5287 ir_mode *mode = get_irn_mode(n);
5289 HANDLE_BINOP_PHI((eval_func) tarval_shl, a, b, c, mode);
5290 n = transform_node_shift(n);
5293 n = transform_node_shift_modulo(n, new_rd_Shl);
5295 n = transform_node_shl_shr(n);
5297 n = transform_node_bitop_shift(n);
5300 } /* transform_node_Shl */
5305 static ir_node *transform_node_Rotl(ir_node *n)
5307 ir_node *c, *oldn = n;
5308 ir_node *a = get_Rotl_left(n);
5309 ir_node *b = get_Rotl_right(n);
5310 ir_mode *mode = get_irn_mode(n);
5312 HANDLE_BINOP_PHI((eval_func) tarval_rotl, a, b, c, mode);
5313 n = transform_node_shift(n);
5316 n = transform_node_bitop_shift(n);
5319 } /* transform_node_Rotl */
5324 static ir_node *transform_node_Conv(ir_node *n)
5326 ir_node *c, *oldn = n;
5327 ir_mode *mode = get_irn_mode(n);
5328 ir_node *a = get_Conv_op(n);
5330 if (mode != mode_b && is_const_Phi(a)) {
5331 /* Do NOT optimize mode_b Conv's, this leads to remaining
5332 * Phib nodes later, because the conv_b_lower operation
5333 * is instantly reverted, when it tries to insert a Convb.
5335 c = apply_conv_on_phi(a, mode);
5337 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5342 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5343 ir_graph *irg = get_irn_irg(n);
5344 return new_r_Unknown(irg, mode);
5347 if (mode_is_reference(mode) &&
5348 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5350 ir_node *l = get_Add_left(a);
5351 ir_node *r = get_Add_right(a);
5352 dbg_info *dbgi = get_irn_dbg_info(a);
5353 ir_node *block = get_nodes_block(n);
5355 ir_node *lop = get_Conv_op(l);
5356 if (get_irn_mode(lop) == mode) {
5357 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5358 n = new_rd_Add(dbgi, block, lop, r, mode);
5363 ir_node *rop = get_Conv_op(r);
5364 if (get_irn_mode(rop) == mode) {
5365 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5366 n = new_rd_Add(dbgi, block, l, rop, mode);
5373 } /* transform_node_Conv */
5376 * Remove dead blocks and nodes in dead blocks
5377 * in keep alive list. We do not generate a new End node.
5379 static ir_node *transform_node_End(ir_node *n)
5381 int i, j, n_keepalives = get_End_n_keepalives(n);
5384 NEW_ARR_A(ir_node *, in, n_keepalives);
5386 for (i = j = 0; i < n_keepalives; ++i) {
5387 ir_node *ka = get_End_keepalive(n, i);
5389 if (! is_Block_dead(ka)) {
5393 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
5395 } else if (is_Bad(ka)) {
5396 /* no need to keep Bad */
5401 if (j != n_keepalives)
5402 set_End_keepalives(n, j, in);
5404 } /* transform_node_End */
5406 int ir_is_negated_value(const ir_node *a, const ir_node *b)
5408 if (is_Minus(a) && get_Minus_op(a) == b)
5410 if (is_Minus(b) && get_Minus_op(b) == a)
5412 if (is_Sub(a) && is_Sub(b)) {
5413 ir_node *a_left = get_Sub_left(a);
5414 ir_node *a_right = get_Sub_right(a);
5415 ir_node *b_left = get_Sub_left(b);
5416 ir_node *b_right = get_Sub_right(b);
5418 if (a_left == b_right && a_right == b_left)
5426 * Optimize a Mux into some simpler cases.
5428 static ir_node *transform_node_Mux(ir_node *n)
5430 ir_node *oldn = n, *sel = get_Mux_sel(n);
5431 ir_mode *mode = get_irn_mode(n);
5432 ir_node *t = get_Mux_true(n);
5433 ir_node *f = get_Mux_false(n);
5434 ir_graph *irg = get_irn_irg(n);
5436 if (is_irg_state(irg, IR_GRAPH_STATE_KEEP_MUX))
5440 ir_node* block = get_nodes_block(n);
5442 ir_node* c1 = get_Mux_sel(t);
5443 ir_node* t1 = get_Mux_true(t);
5444 ir_node* f1 = get_Mux_false(t);
5446 /* Mux(cond0, Mux(cond1, x, y), y) -> typical if (cond0 && cond1) x else y */
5447 ir_node* and_ = new_r_And(block, c0, c1, mode_b);
5448 ir_node* new_mux = new_r_Mux(block, and_, f1, t1, mode);
5453 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5454 } else if (f == t1) {
5455 /* Mux(cond0, Mux(cond1, x, y), x) */
5456 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5457 ir_node* and_ = new_r_And(block, c0, not_c1, mode_b);
5458 ir_node* new_mux = new_r_Mux(block, and_, t1, f1, mode);
5463 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5465 } else if (is_Mux(f)) {
5466 ir_node* block = get_nodes_block(n);
5468 ir_node* c1 = get_Mux_sel(f);
5469 ir_node* t1 = get_Mux_true(f);
5470 ir_node* f1 = get_Mux_false(f);
5472 /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */
5473 ir_node* or_ = new_r_Or(block, c0, c1, mode_b);
5474 ir_node* new_mux = new_r_Mux(block, or_, f1, t1, mode);
5479 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5480 } else if (t == f1) {
5481 /* Mux(cond0, x, Mux(cond1, y, x)) */
5482 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5483 ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b);
5484 ir_node* new_mux = new_r_Mux(block, or_, t1, f1, mode);
5489 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5493 /* first normalization step: try to move a constant to the false side,
5494 * 0 preferred on false side too */
5495 if (is_Cmp(sel) && is_Const(t) &&
5496 (!is_Const(f) || (is_Const_null(t) && !is_Const_null(f)))) {
5497 dbg_info *seldbgi = get_irn_dbg_info(sel);
5498 ir_node *block = get_nodes_block(sel);
5499 ir_relation relation = get_Cmp_relation(sel);
5504 /* Mux(x, a, b) => Mux(not(x), b, a) */
5505 relation = get_negated_relation(relation);
5506 sel = new_rd_Cmp(seldbgi, block, get_Cmp_left(sel),
5507 get_Cmp_right(sel), relation);
5508 n = new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, f, t, mode);
5511 /* note: after normalization, false can only happen on default */
5512 if (mode == mode_b) {
5513 dbg_info *dbg = get_irn_dbg_info(n);
5514 ir_node *block = get_nodes_block(n);
5517 ir_tarval *tv_t = get_Const_tarval(t);
5518 if (tv_t == tarval_b_true) {
5520 /* Muxb(sel, true, false) = sel */
5521 assert(get_Const_tarval(f) == tarval_b_false);
5522 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5525 /* Muxb(sel, true, x) = Or(sel, x) */
5526 n = new_rd_Or(dbg, block, sel, f, mode_b);
5527 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5531 } else if (is_Const(f)) {
5532 ir_tarval *tv_f = get_Const_tarval(f);
5533 if (tv_f == tarval_b_true) {
5534 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5535 ir_node* not_sel = new_rd_Not(dbg, block, sel, mode_b);
5536 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5537 n = new_rd_Or(dbg, block, not_sel, t, mode_b);
5540 /* Muxb(sel, x, false) = And(sel, x) */
5541 assert(tv_f == tarval_b_false);
5542 n = new_rd_And(dbg, block, sel, t, mode_b);
5543 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5549 /* more normalization: Mux(sel, 0, 1) is simply a conv from the mode_b
5550 * value to integer. */
5551 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5552 ir_tarval *a = get_Const_tarval(t);
5553 ir_tarval *b = get_Const_tarval(f);
5555 if (tarval_is_one(a) && tarval_is_null(b)) {
5556 ir_node *block = get_nodes_block(n);
5557 ir_node *conv = new_r_Conv(block, sel, mode);
5559 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5561 } else if (tarval_is_null(a) && tarval_is_one(b)) {
5562 ir_node *block = get_nodes_block(n);
5563 ir_node *not_ = new_r_Not(block, sel, mode_b);
5564 ir_node *conv = new_r_Conv(block, not_, mode);
5566 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5572 ir_node *cmp_r = get_Cmp_right(sel);
5573 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5574 ir_node *block = get_nodes_block(n);
5575 ir_node *cmp_l = get_Cmp_left(sel);
5577 if (mode_is_int(mode)) {
5578 ir_relation relation = get_Cmp_relation(sel);
5580 if ((relation == ir_relation_less_greater || relation == ir_relation_equal) && is_And(cmp_l)) {
5581 /* Mux((a & b) != 0, c, 0) */
5582 ir_node *and_r = get_And_right(cmp_l);
5585 if (and_r == t && f == cmp_r) {
5586 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5587 if (relation == ir_relation_less_greater) {
5588 /* Mux((a & 2^C) != 0, 2^C, 0) */
5590 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5592 /* Mux((a & 2^C) == 0, 2^C, 0) */
5593 n = new_rd_Eor(get_irn_dbg_info(n),
5594 block, cmp_l, t, mode);
5595 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5600 if (is_Shl(and_r)) {
5601 ir_node *shl_l = get_Shl_left(and_r);
5602 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5603 if (and_r == t && f == cmp_r) {
5604 if (relation == ir_relation_less_greater) {
5605 /* (a & (1 << n)) != 0, (1 << n), 0) */
5607 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5609 /* (a & (1 << n)) == 0, (1 << n), 0) */
5610 n = new_rd_Eor(get_irn_dbg_info(n),
5611 block, cmp_l, t, mode);
5612 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5618 and_l = get_And_left(cmp_l);
5619 if (is_Shl(and_l)) {
5620 ir_node *shl_l = get_Shl_left(and_l);
5621 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5622 if (and_l == t && f == cmp_r) {
5623 if (relation == ir_relation_less_greater) {
5624 /* ((1 << n) & a) != 0, (1 << n), 0) */
5626 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5628 /* ((1 << n) & a) == 0, (1 << n), 0) */
5629 n = new_rd_Eor(get_irn_dbg_info(n),
5630 block, cmp_l, t, mode);
5631 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5646 * optimize Sync nodes that have other syncs as input we simply add the inputs
5647 * of the other sync to our own inputs
5649 static ir_node *transform_node_Sync(ir_node *n)
5651 int arity = get_Sync_n_preds(n);
5654 for (i = 0; i < arity;) {
5655 ir_node *pred = get_Sync_pred(n, i);
5659 if (!is_Sync(pred)) {
5667 pred_arity = get_Sync_n_preds(pred);
5668 for (j = 0; j < pred_arity; ++j) {
5669 ir_node *pred_pred = get_Sync_pred(pred, j);
5674 add_irn_n(n, pred_pred);
5678 if (get_Sync_pred(n, k) == pred_pred) break;
5683 /* rehash the sync node */
5687 } /* transform_node_Sync */
5690 * optimize a trampoline Call into a direct Call
5692 static ir_node *transform_node_Call(ir_node *call)
5694 ir_node *callee = get_Call_ptr(call);
5695 ir_node *adr, *mem, *res, *bl, **in;
5696 ir_type *ctp, *mtp, *tp;
5700 size_t i, n_res, n_param;
5703 if (! is_Proj(callee))
5705 callee = get_Proj_pred(callee);
5706 if (! is_Builtin(callee))
5708 if (get_Builtin_kind(callee) != ir_bk_inner_trampoline)
5711 mem = get_Call_mem(call);
5713 if (skip_Proj(mem) == callee) {
5714 /* memory is routed to the trampoline, skip */
5715 mem = get_Builtin_mem(callee);
5718 /* build a new call type */
5719 mtp = get_Call_type(call);
5720 tdb = get_type_dbg_info(mtp);
5722 n_res = get_method_n_ress(mtp);
5723 n_param = get_method_n_params(mtp);
5724 ctp = new_d_type_method(n_param + 1, n_res, tdb);
5726 for (i = 0; i < n_res; ++i)
5727 set_method_res_type(ctp, i, get_method_res_type(mtp, i));
5729 NEW_ARR_A(ir_node *, in, n_param + 1);
5731 /* FIXME: we don't need a new pointer type in every step */
5732 irg = get_irn_irg(call);
5733 tp = get_irg_frame_type(irg);
5734 tp = new_type_pointer(tp);
5735 set_method_param_type(ctp, 0, tp);
5737 in[0] = get_Builtin_param(callee, 2);
5738 for (i = 0; i < n_param; ++i) {
5739 set_method_param_type(ctp, i + 1, get_method_param_type(mtp, i));
5740 in[i + 1] = get_Call_param(call, i);
5742 var = get_method_variadicity(mtp);
5743 set_method_variadicity(ctp, var);
5744 /* When we resolve a trampoline, the function must be called by a this-call */
5745 set_method_calling_convention(ctp, get_method_calling_convention(mtp) | cc_this_call);
5746 set_method_additional_properties(ctp, get_method_additional_properties(mtp));
5748 adr = get_Builtin_param(callee, 1);
5750 db = get_irn_dbg_info(call);
5751 bl = get_nodes_block(call);
5753 res = new_rd_Call(db, bl, mem, adr, n_param + 1, in, ctp);
5754 if (get_irn_pinned(call) == op_pin_state_floats)
5755 set_irn_pinned(res, op_pin_state_floats);
5757 } /* transform_node_Call */
5760 * Tries several [inplace] [optimizing] transformations and returns an
5761 * equivalent node. The difference to equivalent_node() is that these
5762 * transformations _do_ generate new nodes, and thus the old node must
5763 * not be freed even if the equivalent node isn't the old one.
5765 static ir_node *transform_node(ir_node *n)
5770 * Transform_node is the only "optimizing transformation" that might
5771 * return a node with a different opcode. We iterate HERE until fixpoint
5772 * to get the final result.
5776 if (n->op->ops.transform_node != NULL)
5777 n = n->op->ops.transform_node(n);
5778 } while (oldn != n);
5781 } /* transform_node */
5784 * Sets the default transform node operation for an ir_op_ops.
5786 * @param code the opcode for the default operation
5787 * @param ops the operations initialized
5792 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5796 ops->transform_node = transform_node_##a; \
5798 #define CASE_PROJ(a) \
5800 ops->transform_node_Proj = transform_node_Proj_##a; \
5802 #define CASE_PROJ_EX(a) \
5804 ops->transform_node = transform_node_##a; \
5805 ops->transform_node_Proj = transform_node_Proj_##a; \
5845 } /* firm_set_default_transform_node */
5848 /* **************** Common Subexpression Elimination **************** */
5850 /** The size of the hash table used, should estimate the number of nodes
5852 #define N_IR_NODES 512
5854 /** Compares the attributes of two Const nodes. */
5855 static int node_cmp_attr_Const(const ir_node *a, const ir_node *b)
5857 return get_Const_tarval(a) != get_Const_tarval(b);
5860 /** Compares the attributes of two Proj nodes. */
5861 static int node_cmp_attr_Proj(const ir_node *a, const ir_node *b)
5863 return a->attr.proj.proj != b->attr.proj.proj;
5866 /** Compares the attributes of two Alloc nodes. */
5867 static int node_cmp_attr_Alloc(const ir_node *a, const ir_node *b)
5869 const alloc_attr *pa = &a->attr.alloc;
5870 const alloc_attr *pb = &b->attr.alloc;
5871 return (pa->where != pb->where) || (pa->type != pb->type);
5874 /** Compares the attributes of two Free nodes. */
5875 static int node_cmp_attr_Free(const ir_node *a, const ir_node *b)
5877 const free_attr *pa = &a->attr.free;
5878 const free_attr *pb = &b->attr.free;
5879 return (pa->where != pb->where) || (pa->type != pb->type);
5882 /** Compares the attributes of two SymConst nodes. */
5883 static int node_cmp_attr_SymConst(const ir_node *a, const ir_node *b)
5885 const symconst_attr *pa = &a->attr.symc;
5886 const symconst_attr *pb = &b->attr.symc;
5887 return (pa->kind != pb->kind)
5888 || (pa->sym.type_p != pb->sym.type_p);
5891 /** Compares the attributes of two Call nodes. */
5892 static int node_cmp_attr_Call(const ir_node *a, const ir_node *b)
5894 const call_attr *pa = &a->attr.call;
5895 const call_attr *pb = &b->attr.call;
5896 return (pa->type != pb->type)
5897 || (pa->tail_call != pb->tail_call);
5900 /** Compares the attributes of two Sel nodes. */
5901 static int node_cmp_attr_Sel(const ir_node *a, const ir_node *b)
5903 const ir_entity *a_ent = get_Sel_entity(a);
5904 const ir_entity *b_ent = get_Sel_entity(b);
5905 return a_ent != b_ent;
5908 /** Compares the attributes of two Phi nodes. */
5909 static int node_cmp_attr_Phi(const ir_node *a, const ir_node *b)
5911 /* we can only enter this function if both nodes have the same number of inputs,
5912 hence it is enough to check if one of them is a Phi0 */
5914 /* check the Phi0 pos attribute */
5915 return a->attr.phi.u.pos != b->attr.phi.u.pos;
5920 /** Compares the attributes of two Conv nodes. */
5921 static int node_cmp_attr_Conv(const ir_node *a, const ir_node *b)
5923 return get_Conv_strict(a) != get_Conv_strict(b);
5926 /** Compares the attributes of two Cast nodes. */
5927 static int node_cmp_attr_Cast(const ir_node *a, const ir_node *b)
5929 return get_Cast_type(a) != get_Cast_type(b);
5932 /** Compares the attributes of two Load nodes. */
5933 static int node_cmp_attr_Load(const ir_node *a, const ir_node *b)
5935 if (get_Load_volatility(a) == volatility_is_volatile ||
5936 get_Load_volatility(b) == volatility_is_volatile)
5937 /* NEVER do CSE on volatile Loads */
5939 /* do not CSE Loads with different alignment. Be conservative. */
5940 if (get_Load_unaligned(a) != get_Load_unaligned(b))
5943 return get_Load_mode(a) != get_Load_mode(b);
5946 /** Compares the attributes of two Store nodes. */
5947 static int node_cmp_attr_Store(const ir_node *a, const ir_node *b)
5949 /* do not CSE Stores with different alignment. Be conservative. */
5950 if (get_Store_unaligned(a) != get_Store_unaligned(b))
5953 /* NEVER do CSE on volatile Stores */
5954 return (get_Store_volatility(a) == volatility_is_volatile ||
5955 get_Store_volatility(b) == volatility_is_volatile);
5958 /** Compares two exception attributes */
5959 static int node_cmp_exception(const ir_node *a, const ir_node *b)
5961 const except_attr *ea = &a->attr.except;
5962 const except_attr *eb = &b->attr.except;
5964 return ea->pin_state != eb->pin_state;
5967 #define node_cmp_attr_Bound node_cmp_exception
5969 /** Compares the attributes of two Div nodes. */
5970 static int node_cmp_attr_Div(const ir_node *a, const ir_node *b)
5972 const div_attr *ma = &a->attr.div;
5973 const div_attr *mb = &b->attr.div;
5974 return ma->exc.pin_state != mb->exc.pin_state ||
5975 ma->resmode != mb->resmode ||
5976 ma->no_remainder != mb->no_remainder;
5979 /** Compares the attributes of two Mod nodes. */
5980 static int node_cmp_attr_Mod(const ir_node *a, const ir_node *b)
5982 const mod_attr *ma = &a->attr.mod;
5983 const mod_attr *mb = &b->attr.mod;
5984 return ma->exc.pin_state != mb->exc.pin_state ||
5985 ma->resmode != mb->resmode;
5988 static int node_cmp_attr_Cmp(const ir_node *a, const ir_node *b)
5990 const cmp_attr *ma = &a->attr.cmp;
5991 const cmp_attr *mb = &b->attr.cmp;
5992 return ma->relation != mb->relation;
5995 /** Compares the attributes of two Confirm nodes. */
5996 static int node_cmp_attr_Confirm(const ir_node *a, const ir_node *b)
5998 const confirm_attr *ma = &a->attr.confirm;
5999 const confirm_attr *mb = &b->attr.confirm;
6000 return ma->relation != mb->relation;
6003 /** Compares the attributes of two Builtin nodes. */
6004 static int node_cmp_attr_Builtin(const ir_node *a, const ir_node *b)
6006 /* no need to compare the type, equal kind means equal type */
6007 return get_Builtin_kind(a) != get_Builtin_kind(b);
6010 /** Compares the attributes of two ASM nodes. */
6011 static int node_cmp_attr_ASM(const ir_node *a, const ir_node *b)
6014 const ir_asm_constraint *ca;
6015 const ir_asm_constraint *cb;
6018 if (get_ASM_text(a) != get_ASM_text(b))
6021 /* Should we really check the constraints here? Should be better, but is strange. */
6022 n = get_ASM_n_input_constraints(a);
6023 if (n != get_ASM_n_input_constraints(b))
6026 ca = get_ASM_input_constraints(a);
6027 cb = get_ASM_input_constraints(b);
6028 for (i = 0; i < n; ++i) {
6029 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
6033 n = get_ASM_n_output_constraints(a);
6034 if (n != get_ASM_n_output_constraints(b))
6037 ca = get_ASM_output_constraints(a);
6038 cb = get_ASM_output_constraints(b);
6039 for (i = 0; i < n; ++i) {
6040 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
6044 n = get_ASM_n_clobbers(a);
6045 if (n != get_ASM_n_clobbers(b))
6048 cla = get_ASM_clobbers(a);
6049 clb = get_ASM_clobbers(b);
6050 for (i = 0; i < n; ++i) {
6051 if (cla[i] != clb[i])
6057 /** Compares the inexistent attributes of two Dummy nodes. */
6058 static int node_cmp_attr_Dummy(const ir_node *a, const ir_node *b)
6066 * Set the default node attribute compare operation for an ir_op_ops.
6068 * @param code the opcode for the default operation
6069 * @param ops the operations initialized
6074 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
6078 ops->node_cmp_attr = node_cmp_attr_##a; \
6110 } /* firm_set_default_node_cmp_attr */
6113 * Compare function for two nodes in the value table. Gets two
6114 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6116 int identities_cmp(const void *elt, const void *key)
6118 ir_node *a = (ir_node *)elt;
6119 ir_node *b = (ir_node *)key;
6122 if (a == b) return 0;
6124 if ((get_irn_op(a) != get_irn_op(b)) ||
6125 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6127 /* compare if a's in and b's in are of equal length */
6128 irn_arity_a = get_irn_arity(a);
6129 if (irn_arity_a != get_irn_arity(b))
6132 /* blocks are never the same */
6136 if (get_irn_pinned(a) == op_pin_state_pinned) {
6137 /* for pinned nodes, the block inputs must be equal */
6138 if (get_irn_n(a, -1) != get_irn_n(b, -1))
6140 } else if (! get_opt_global_cse()) {
6141 /* for block-local CSE both nodes must be in the same Block */
6142 if (get_nodes_block(a) != get_nodes_block(b))
6146 /* compare a->in[0..ins] with b->in[0..ins] */
6147 for (i = 0; i < irn_arity_a; ++i) {
6148 ir_node *pred_a = get_irn_n(a, i);
6149 ir_node *pred_b = get_irn_n(b, i);
6150 if (pred_a != pred_b) {
6151 /* if both predecessors are CSE neutral they might be different */
6152 if (!is_irn_cse_neutral(pred_a) || !is_irn_cse_neutral(pred_b))
6158 * here, we already now that the nodes are identical except their
6161 if (a->op->ops.node_cmp_attr)
6162 return a->op->ops.node_cmp_attr(a, b);
6165 } /* identities_cmp */
6168 * Calculate a hash value of a node.
6170 * @param node The IR-node
6172 unsigned ir_node_hash(const ir_node *node)
6174 return node->op->ops.hash(node);
6175 } /* ir_node_hash */
6178 void new_identities(ir_graph *irg)
6180 if (irg->value_table != NULL)
6181 del_pset(irg->value_table);
6182 irg->value_table = new_pset(identities_cmp, N_IR_NODES);
6183 } /* new_identities */
6185 void del_identities(ir_graph *irg)
6187 if (irg->value_table != NULL)
6188 del_pset(irg->value_table);
6189 } /* del_identities */
6191 /* Normalize a node by putting constants (and operands with larger
6192 * node index) on the right (operator side). */
6193 void ir_normalize_node(ir_node *n)
6195 if (is_op_commutative(get_irn_op(n))) {
6196 ir_node *l = get_binop_left(n);
6197 ir_node *r = get_binop_right(n);
6199 /* For commutative operators perform a OP b == b OP a but keep
6200 * constants on the RIGHT side. This helps greatly in some
6201 * optimizations. Moreover we use the idx number to make the form
6203 if (!operands_are_normalized(l, r)) {
6204 set_binop_left(n, r);
6205 set_binop_right(n, l);
6209 } /* ir_normalize_node */
6212 * Return the canonical node computing the same value as n.
6213 * Looks up the node in a hash table, enters it in the table
6214 * if it isn't there yet.
6216 * @param n the node to look up
6218 * @return a node that computes the same value as n or n if no such
6219 * node could be found
6221 ir_node *identify_remember(ir_node *n)
6223 ir_graph *irg = get_irn_irg(n);
6224 pset *value_table = irg->value_table;
6227 if (value_table == NULL)
6230 ir_normalize_node(n);
6231 /* lookup or insert in hash table with given hash key. */
6232 nn = (ir_node*)pset_insert(value_table, n, ir_node_hash(n));
6235 /* n is reachable again */
6236 edges_node_revival(nn);
6240 } /* identify_remember */
6243 * During construction we set the op_pin_state_pinned flag in the graph right
6244 * when the optimization is performed. The flag turning on procedure global
6245 * cse could be changed between two allocations. This way we are safe.
6247 * @param n The node to lookup
6249 static inline ir_node *identify_cons(ir_node *n)
6253 n = identify_remember(n);
6254 if (n != old && get_nodes_block(old) != get_nodes_block(n)) {
6255 ir_graph *irg = get_irn_irg(n);
6256 set_irg_pinned(irg, op_pin_state_floats);
6259 } /* identify_cons */
6261 /* Add a node to the identities value table. */
6262 void add_identities(ir_node *node)
6269 identify_remember(node);
6272 /* Visit each node in the value table of a graph. */
6273 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env)
6276 ir_graph *rem = current_ir_graph;
6278 current_ir_graph = irg;
6279 foreach_pset(irg->value_table, ir_node*, node) {
6282 current_ir_graph = rem;
6283 } /* visit_all_identities */
6286 * Garbage in, garbage out. If a node has a dead input, i.e., the
6287 * Bad node is input to the node, return the Bad node.
6289 static ir_node *gigo(ir_node *node)
6292 ir_op *op = get_irn_op(node);
6294 /* remove garbage blocks by looking at control flow that leaves the block
6295 and replacing the control flow by Bad. */
6296 if (get_irn_mode(node) == mode_X) {
6297 ir_node *block = get_nodes_block(skip_Proj(node));
6298 ir_graph *irg = get_irn_irg(block);
6300 /* Don't optimize nodes in immature blocks. */
6301 if (!get_Block_matured(block))
6303 /* Don't optimize End, may have Bads. */
6304 if (op == op_End) return node;
6306 if (is_Block(block)) {
6307 if (is_Block_dead(block)) {
6308 /* control flow from dead block is dead */
6309 return get_irg_bad(irg);
6312 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
6313 if (!is_Bad(get_irn_n(block, i)))
6317 ir_graph *irg = get_irn_irg(block);
6318 /* the start block is never dead */
6319 if (block != get_irg_start_block(irg)
6320 && block != get_irg_end_block(irg)) {
6322 * Do NOT kill control flow without setting
6323 * the block to dead of bad things can happen:
6324 * We get a Block that is not reachable be irg_block_walk()
6325 * but can be found by irg_walk()!
6327 set_Block_dead(block);
6328 return get_irg_bad(irg);
6334 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6335 blocks predecessors is dead. */
6336 if (op != op_Block && op != op_Phi && op != op_Tuple && op != op_Anchor) {
6337 ir_graph *irg = get_irn_irg(node);
6338 irn_arity = get_irn_arity(node);
6341 * Beware: we can only read the block of a non-floating node.
6343 if (is_irn_pinned_in_irg(node) &&
6344 is_Block_dead(get_nodes_block(skip_Proj(node))))
6345 return get_irg_bad(irg);
6347 for (i = 0; i < irn_arity; i++) {
6348 ir_node *pred = get_irn_n(node, i);
6351 return get_irg_bad(irg);
6353 /* Propagating Unknowns here seems to be a bad idea, because
6354 sometimes we need a node as a input and did not want that
6356 However, it might be useful to move this into a later phase
6357 (if you think that optimizing such code is useful). */
6358 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
6359 return new_r_Unknown(irg, get_irn_mode(node));
6364 /* With this code we violate the agreement that local_optimize
6365 only leaves Bads in Block, Phi and Tuple nodes. */
6366 /* If Block has only Bads as predecessors it's garbage. */
6367 /* If Phi has only Bads as predecessors it's garbage. */
6368 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
6369 irn_arity = get_irn_arity(node);
6370 for (i = 0; i < irn_arity; i++) {
6371 if (!is_Bad(get_irn_n(node, i))) break;
6373 if (i == irn_arity) node = get_irg_bad(irg);
6380 * These optimizations deallocate nodes from the obstack.
6381 * It can only be called if it is guaranteed that no other nodes
6382 * reference this one, i.e., right after construction of a node.
6384 * @param n The node to optimize
6386 ir_node *optimize_node(ir_node *n)
6389 ir_graph *irg = get_irn_irg(n);
6390 unsigned iro = get_irn_opcode(n);
6393 /* Always optimize Phi nodes: part of the construction. */
6394 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6396 /* Remove nodes with dead (Bad) input.
6397 Run always for transformation induced Bads. */
6400 edges_node_deleted(oldn);
6402 /* We found an existing, better node, so we can deallocate the old node. */
6403 irg_kill_node(irg, oldn);
6407 /* constant expression evaluation / constant folding */
6408 if (get_opt_constant_folding()) {
6409 /* neither constants nor Tuple values can be evaluated */
6410 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6411 /* try to evaluate */
6412 tv = computed_value(n);
6413 if (tv != tarval_bad) {
6418 * we MUST copy the node here temporarily, because it's still
6419 * needed for DBG_OPT_CSTEVAL
6421 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6422 oldn = (ir_node*)alloca(node_size);
6424 memcpy(oldn, n, node_size);
6425 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6427 /* ARG, copy the in array, we need it for statistics */
6428 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6430 /* note the inplace edges module */
6431 edges_node_deleted(n);
6433 /* evaluation was successful -- replace the node. */
6434 irg_kill_node(irg, n);
6435 nw = new_r_Const(irg, tv);
6437 DBG_OPT_CSTEVAL(oldn, nw);
6443 /* remove unnecessary nodes */
6444 if (get_opt_algebraic_simplification() ||
6445 (iro == iro_Phi) || /* always optimize these nodes. */
6447 (iro == iro_Proj) ||
6448 (iro == iro_Block) ) /* Flags tested local. */
6449 n = equivalent_node(n);
6451 /* Common Subexpression Elimination.
6453 * Checks whether n is already available.
6454 * The block input is used to distinguish different subexpressions. Right
6455 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6456 * subexpressions within a block.
6459 n = identify_cons(n);
6462 edges_node_deleted(oldn);
6464 /* We found an existing, better node, so we can deallocate the old node. */
6465 irg_kill_node(irg, oldn);
6469 /* Some more constant expression evaluation that does not allow to
6471 iro = get_irn_opcode(n);
6472 if (get_opt_algebraic_simplification() ||
6473 (iro == iro_Cond) ||
6474 (iro == iro_Proj)) /* Flags tested local. */
6475 n = transform_node(n);
6477 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6478 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6480 n = identify_remember(o);
6486 } /* optimize_node */
6490 * These optimizations never deallocate nodes (in place). This can cause dead
6491 * nodes lying on the obstack. Remove these by a dead node elimination,
6492 * i.e., a copying garbage collection.
6494 ir_node *optimize_in_place_2(ir_node *n)
6498 unsigned iro = get_irn_opcode(n);
6500 if (!get_opt_optimize() && !is_Phi(n)) return n;
6502 if (iro == iro_Deleted)
6505 /* Remove nodes with dead (Bad) input.
6506 Run always for transformation induced Bads. */
6511 /* constant expression evaluation / constant folding */
6512 if (get_opt_constant_folding()) {
6513 /* neither constants nor Tuple values can be evaluated */
6514 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6515 /* try to evaluate */
6516 tv = computed_value(n);
6517 if (tv != tarval_bad) {
6518 /* evaluation was successful -- replace the node. */
6519 ir_graph *irg = get_irn_irg(n);
6521 n = new_r_Const(irg, tv);
6523 DBG_OPT_CSTEVAL(oldn, n);
6529 /* remove unnecessary nodes */
6530 if (get_opt_constant_folding() ||
6531 (iro == iro_Phi) || /* always optimize these nodes. */
6532 (iro == iro_Id) || /* ... */
6533 (iro == iro_Proj) || /* ... */
6534 (iro == iro_Block) ) /* Flags tested local. */
6535 n = equivalent_node(n);
6537 /** common subexpression elimination **/
6538 /* Checks whether n is already available. */
6539 /* The block input is used to distinguish different subexpressions. Right
6540 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6541 subexpressions within a block. */
6542 if (get_opt_cse()) {
6544 n = identify_remember(o);
6549 /* Some more constant expression evaluation. */
6550 iro = get_irn_opcode(n);
6551 if (get_opt_constant_folding() ||
6552 (iro == iro_Cond) ||
6553 (iro == iro_Proj)) /* Flags tested local. */
6554 n = transform_node(n);
6556 /* Now we can verify the node, as it has no dead inputs any more. */
6559 /* Now we have a legal, useful node. Enter it in hash table for cse.
6560 Blocks should be unique anyways. (Except the successor of start:
6561 is cse with the start block!) */
6562 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6564 n = identify_remember(o);
6570 } /* optimize_in_place_2 */
6573 * Wrapper for external use, set proper status bits after optimization.
6575 ir_node *optimize_in_place(ir_node *n)
6577 ir_graph *irg = get_irn_irg(n);
6578 /* Handle graph state */
6579 assert(get_irg_phase_state(irg) != phase_building);
6581 if (get_opt_global_cse())
6582 set_irg_pinned(irg, op_pin_state_floats);
6583 if (get_irg_outs_state(irg) == outs_consistent)
6584 set_irg_outs_inconsistent(irg);
6586 /* FIXME: Maybe we could also test whether optimizing the node can
6587 change the control graph. */
6588 set_irg_doms_inconsistent(irg);
6589 return optimize_in_place_2(n);
6590 } /* optimize_in_place */
6593 * Calculate a hash value of a Const node.
6595 static unsigned hash_Const(const ir_node *node)
6599 /* special value for const, as they only differ in their tarval. */
6600 h = HASH_PTR(node->attr.con.tarval);
6606 * Calculate a hash value of a SymConst node.
6608 static unsigned hash_SymConst(const ir_node *node)
6612 /* all others are pointers */
6613 h = HASH_PTR(node->attr.symc.sym.type_p);
6616 } /* hash_SymConst */
6619 * Set the default hash operation in an ir_op_ops.
6621 * @param code the opcode for the default operation
6622 * @param ops the operations initialized
6627 static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops)
6631 ops->hash = hash_##a; \
6634 /* hash function already set */
6635 if (ops->hash != NULL)
6642 /* use input/mode default hash if no function was given */
6643 ops->hash = firm_default_hash;
6651 * Sets the default operation for an ir_ops.
6653 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops)
6655 ops = firm_set_default_hash(code, ops);
6656 ops = firm_set_default_computed_value(code, ops);
6657 ops = firm_set_default_equivalent_node(code, ops);
6658 ops = firm_set_default_transform_node(code, ops);
6659 ops = firm_set_default_node_cmp_attr(code, ops);
6660 ops = firm_set_default_get_type_attr(code, ops);
6661 ops = firm_set_default_get_entity_attr(code, ops);
6664 } /* firm_set_default_operations */