2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief iropt --- optimizations intertwined with IR construction.
23 * @author Christian Schaefer, Goetz Lindenmaier, Michael Beck
31 #include "irgraph_t.h"
32 #include "iredges_t.h"
39 #include "dbginfo_t.h"
40 #include "iropt_dbg.h"
45 #include "opt_confirms.h"
46 #include "opt_polymorphy.h"
51 #include "firm_types.h"
53 /* Make types visible to allow most efficient access */
57 * Returns the tarval of a Const node or tarval_bad for all other nodes.
59 static tarval *default_value_of(const ir_node *n)
62 return get_Const_tarval(n); /* might return tarval_bad */
67 value_of_func value_of_ptr = default_value_of;
69 /* * Set a new value_of function. */
70 void set_value_of_func(value_of_func func)
75 value_of_ptr = default_value_of;
79 * Return the value of a Constant.
81 static tarval *computed_value_Const(const ir_node *n)
83 return get_Const_tarval(n);
84 } /* computed_value_Const */
87 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
89 static tarval *computed_value_SymConst(const ir_node *n)
94 switch (get_SymConst_kind(n)) {
95 case symconst_type_size:
96 type = get_SymConst_type(n);
97 if (get_type_state(type) == layout_fixed)
98 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
100 case symconst_type_align:
101 type = get_SymConst_type(n);
102 if (get_type_state(type) == layout_fixed)
103 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
105 case symconst_ofs_ent:
106 ent = get_SymConst_entity(n);
107 type = get_entity_owner(ent);
108 if (get_type_state(type) == layout_fixed)
109 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
115 } /* computed_value_SymConst */
118 * Return the value of an Add.
120 static tarval *computed_value_Add(const ir_node *n)
122 ir_node *a = get_Add_left(n);
123 ir_node *b = get_Add_right(n);
125 tarval *ta = value_of(a);
126 tarval *tb = value_of(b);
128 if ((ta != tarval_bad) && (tb != tarval_bad))
129 return tarval_add(ta, tb);
132 } /* computed_value_Add */
135 * Return the value of a Sub.
136 * Special case: a - a
138 static tarval *computed_value_Sub(const ir_node *n)
140 ir_mode *mode = get_irn_mode(n);
141 ir_node *a = get_Sub_left(n);
142 ir_node *b = get_Sub_right(n);
147 if (! mode_is_float(mode)) {
150 return get_mode_null(mode);
156 if ((ta != tarval_bad) && (tb != tarval_bad))
157 return tarval_sub(ta, tb, mode);
160 } /* computed_value_Sub */
163 * Return the value of a Carry.
164 * Special : a op 0, 0 op b
166 static tarval *computed_value_Carry(const ir_node *n)
168 ir_node *a = get_binop_left(n);
169 ir_node *b = get_binop_right(n);
170 ir_mode *m = get_irn_mode(n);
172 tarval *ta = value_of(a);
173 tarval *tb = value_of(b);
175 if ((ta != tarval_bad) && (tb != tarval_bad)) {
177 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
179 if (tarval_is_null(ta) || tarval_is_null(tb))
180 return get_mode_null(m);
183 } /* computed_value_Carry */
186 * Return the value of a Borrow.
189 static tarval *computed_value_Borrow(const ir_node *n)
191 ir_node *a = get_binop_left(n);
192 ir_node *b = get_binop_right(n);
193 ir_mode *m = get_irn_mode(n);
195 tarval *ta = value_of(a);
196 tarval *tb = value_of(b);
198 if ((ta != tarval_bad) && (tb != tarval_bad)) {
199 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
200 } else if (tarval_is_null(ta)) {
201 return get_mode_null(m);
204 } /* computed_value_Borrow */
207 * Return the value of an unary Minus.
209 static tarval *computed_value_Minus(const ir_node *n)
211 ir_node *a = get_Minus_op(n);
212 tarval *ta = value_of(a);
214 if (ta != tarval_bad)
215 return tarval_neg(ta);
218 } /* computed_value_Minus */
221 * Return the value of a Mul.
223 static tarval *computed_value_Mul(const ir_node *n)
225 ir_node *a = get_Mul_left(n);
226 ir_node *b = get_Mul_right(n);
229 tarval *ta = value_of(a);
230 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 Abs.
257 static tarval *computed_value_Abs(const ir_node *n)
259 ir_node *a = get_Abs_op(n);
260 tarval *ta = value_of(a);
262 if (ta != tarval_bad)
263 return tarval_abs(ta);
266 } /* computed_value_Abs */
269 * Return the value of an And.
270 * Special case: a & 0, 0 & b
272 static tarval *computed_value_And(const ir_node *n)
274 ir_node *a = get_And_left(n);
275 ir_node *b = get_And_right(n);
277 tarval *ta = value_of(a);
278 tarval *tb = value_of(b);
280 if ((ta != tarval_bad) && (tb != tarval_bad)) {
281 return tarval_and (ta, tb);
283 if (tarval_is_null(ta)) return ta;
284 if (tarval_is_null(tb)) return tb;
287 } /* computed_value_And */
290 * Return the value of an Or.
291 * Special case: a | 1...1, 1...1 | b
293 static tarval *computed_value_Or(const ir_node *n)
295 ir_node *a = get_Or_left(n);
296 ir_node *b = get_Or_right(n);
298 tarval *ta = value_of(a);
299 tarval *tb = value_of(b);
301 if ((ta != tarval_bad) && (tb != tarval_bad)) {
302 return tarval_or (ta, tb);
304 if (tarval_is_all_one(ta)) return ta;
305 if (tarval_is_all_one(tb)) return tb;
308 } /* computed_value_Or */
311 * Return the value of an Eor.
313 static tarval *computed_value_Eor(const ir_node *n)
315 ir_node *a = get_Eor_left(n);
316 ir_node *b = get_Eor_right(n);
321 return get_mode_null(get_irn_mode(n));
326 if ((ta != tarval_bad) && (tb != tarval_bad)) {
327 return tarval_eor(ta, tb);
330 } /* computed_value_Eor */
333 * Return the value of a Not.
335 static tarval *computed_value_Not(const ir_node *n)
337 ir_node *a = get_Not_op(n);
338 tarval *ta = value_of(a);
340 if (ta != tarval_bad)
341 return tarval_not(ta);
344 } /* computed_value_Not */
347 * Return the value of a Shl.
349 static tarval *computed_value_Shl(const ir_node *n)
351 ir_node *a = get_Shl_left(n);
352 ir_node *b = get_Shl_right(n);
354 tarval *ta = value_of(a);
355 tarval *tb = value_of(b);
357 if ((ta != tarval_bad) && (tb != tarval_bad)) {
358 return tarval_shl(ta, tb);
361 } /* computed_value_Shl */
364 * Return the value of a Shr.
366 static tarval *computed_value_Shr(const ir_node *n)
368 ir_node *a = get_Shr_left(n);
369 ir_node *b = get_Shr_right(n);
371 tarval *ta = value_of(a);
372 tarval *tb = value_of(b);
374 if ((ta != tarval_bad) && (tb != tarval_bad)) {
375 return tarval_shr(ta, tb);
378 } /* computed_value_Shr */
381 * Return the value of a Shrs.
383 static tarval *computed_value_Shrs(const ir_node *n)
385 ir_node *a = get_Shrs_left(n);
386 ir_node *b = get_Shrs_right(n);
388 tarval *ta = value_of(a);
389 tarval *tb = value_of(b);
391 if ((ta != tarval_bad) && (tb != tarval_bad)) {
392 return tarval_shrs(ta, tb);
395 } /* computed_value_Shrs */
398 * Return the value of a Rotl.
400 static tarval *computed_value_Rotl(const ir_node *n)
402 ir_node *a = get_Rotl_left(n);
403 ir_node *b = get_Rotl_right(n);
405 tarval *ta = value_of(a);
406 tarval *tb = value_of(b);
408 if ((ta != tarval_bad) && (tb != tarval_bad)) {
409 return tarval_rotl(ta, tb);
412 } /* computed_value_Rotl */
415 * Return the value of a Conv.
417 static tarval *computed_value_Conv(const ir_node *n)
419 ir_node *a = get_Conv_op(n);
420 tarval *ta = value_of(a);
422 if (ta != tarval_bad)
423 return tarval_convert_to(ta, get_irn_mode(n));
426 } /* computed_value_Conv */
429 * Calculate the value of a Mux: can be evaluated, if the
430 * sel and the right input are known.
432 static tarval *computed_value_Mux(const ir_node *n)
434 ir_node *sel = get_Mux_sel(n);
435 tarval *ts = value_of(sel);
437 if (ts == get_tarval_b_true()) {
438 ir_node *v = get_Mux_true(n);
441 else if (ts == get_tarval_b_false()) {
442 ir_node *v = get_Mux_false(n);
446 } /* computed_value_Mux */
449 * Calculate the value of a Confirm: can be evaluated,
450 * if it has the form Confirm(x, '=', Const).
452 static tarval *computed_value_Confirm(const ir_node *n)
455 * Beware: we might produce Phi(Confirm(x == true), Confirm(x == false)).
456 * Do NOT optimize them away (jump threading wants them), so wait until
457 * remove_confirm is activated.
459 if (get_opt_remove_confirm()) {
460 if (get_Confirm_cmp(n) == pn_Cmp_Eq) {
461 tarval *tv = value_of(get_Confirm_bound(n));
462 if (tv != tarval_bad)
466 return value_of(get_Confirm_value(n));
467 } /* computed_value_Confirm */
470 * Return the value of a Proj(Cmp).
472 * This performs a first step of unreachable code elimination.
473 * Proj can not be computed, but folding a Cmp above the Proj here is
474 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
476 * There are several case where we can evaluate a Cmp node, see later.
478 static tarval *computed_value_Proj_Cmp(const ir_node *n)
480 ir_node *cmp = get_Proj_pred(n);
481 ir_node *left = get_Cmp_left(cmp);
482 ir_node *right = get_Cmp_right(cmp);
483 long pn_cmp = get_Proj_proj(n);
484 ir_mode *mode = get_irn_mode(left);
488 * BEWARE: a == a is NOT always True for floating Point values, as
489 * NaN != NaN is defined, so we must check this here.
491 if (left == right && (!mode_is_float(mode) || pn_cmp == pn_Cmp_Lt || pn_cmp == pn_Cmp_Gt)) {
492 /* This is a trick with the bits used for encoding the Cmp
493 Proj numbers, the following statement is not the same:
494 return new_tarval_from_long(pn_cmp == pn_Cmp_Eq, mode_b) */
495 return new_tarval_from_long(pn_cmp & pn_Cmp_Eq, mode_b);
497 tv_l = value_of(left);
498 tv_r = value_of(right);
500 if ((tv_l != tarval_bad) && (tv_r != tarval_bad)) {
502 * The predecessors of Cmp are target values. We can evaluate
505 pn_Cmp flags = tarval_cmp(tv_l, tv_r);
506 if (flags != pn_Cmp_False) {
507 return new_tarval_from_long (pn_cmp & flags, mode_b);
509 } else if (mode_is_int(mode)) {
510 /* for integer values, we can check against MIN/MAX */
513 if (tv_l == get_mode_min(mode)) {
514 /* MIN <=/> x. This results in true/false. */
515 if (pn_cmp == pn_Cmp_Le)
516 return tarval_b_true;
517 else if (pn_cmp == pn_Cmp_Gt)
518 return tarval_b_false;
519 } else if (tv_r == get_mode_min(mode)) {
520 /* x >=/< MIN. This results in true/false. */
521 if (pn_cmp == pn_Cmp_Ge)
522 return tarval_b_true;
523 else if (pn_cmp == pn_Cmp_Lt)
524 return tarval_b_false;
525 } else if (tv_l == get_mode_max(mode)) {
526 /* MAX >=/< x. This results in true/false. */
527 if (pn_cmp == pn_Cmp_Ge)
528 return tarval_b_true;
529 else if (pn_cmp == pn_Cmp_Lt)
530 return tarval_b_false;
531 } else if (tv_r == get_mode_max(mode)) {
532 /* x <=/> MAX. This results in true/false. */
533 if (pn_cmp == pn_Cmp_Le)
534 return tarval_b_true;
535 else if (pn_cmp == pn_Cmp_Gt)
536 return tarval_b_false;
539 cmp_result = vrp_cmp(left, right);
540 if (cmp_result != pn_Cmp_False) {
541 if (cmp_result == pn_Cmp_Lg) {
542 if (pn_cmp == pn_Cmp_Eq) {
543 return tarval_b_false;
544 } else if (pn_cmp == pn_Cmp_Lg) {
545 return tarval_b_true;
548 return new_tarval_from_long(cmp_result & pn_cmp, mode_b);
551 } else if (mode_is_reference(mode)) {
552 /* pointer compare */
553 ir_node *s_l = skip_Proj(left);
554 ir_node *s_r = skip_Proj(right);
556 if ((is_Alloc(s_l) && tarval_is_null(tv_r)) ||
557 (tarval_is_null(tv_l) && is_Alloc(s_r))) {
559 * The predecessors are Allocs and (void*)(0) constants. In Firm Allocs never
560 * return NULL, they raise an exception. Therefore we can predict
563 return new_tarval_from_long(pn_cmp & pn_Cmp_Lg, mode_b);
566 return computed_value_Cmp_Confirm(cmp, left, right, pn_cmp);
567 } /* computed_value_Proj_Cmp */
570 * Return the value of a floating point Quot.
572 static tarval *do_computed_value_Quot(const ir_node *a, const ir_node *b)
574 tarval *ta = value_of(a);
575 tarval *tb = value_of(b);
577 /* cannot optimize 0 / b = 0 because of NaN */
578 if (ta != tarval_bad && tb != tarval_bad)
579 return tarval_quo(ta, tb);
581 } /* do_computed_value_Quot */
584 * Calculate the value of an integer Div of two nodes.
585 * Special case: 0 / b
587 static tarval *do_computed_value_Div(const ir_node *a, const ir_node *b)
589 tarval *ta = value_of(a);
591 const ir_node *dummy;
593 /* Compute c1 / c2 or 0 / a, a != 0 */
594 if (tarval_is_null(ta) && value_not_zero(b, &dummy))
595 return ta; /* 0 / b == 0 */
597 if (ta != tarval_bad && tb != tarval_bad)
598 return tarval_div(ta, tb);
600 } /* do_computed_value_Div */
603 * Calculate the value of an integer Mod of two nodes.
604 * Special case: a % 1
606 static tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b)
608 tarval *ta = value_of(a);
609 tarval *tb = value_of(b);
611 /* Compute a % 1 or c1 % c2 */
612 if (tarval_is_one(tb))
613 return get_mode_null(get_irn_mode(a));
614 if (ta != tarval_bad && tb != tarval_bad)
615 return tarval_mod(ta, tb);
617 } /* do_computed_value_Mod */
620 * Return the value of a Proj(DivMod).
622 static tarval *computed_value_Proj_DivMod(const ir_node *n)
624 long proj_nr = get_Proj_proj(n);
626 /* compute either the Div or the Mod part */
627 if (proj_nr == pn_DivMod_res_div) {
628 const ir_node *a = get_Proj_pred(n);
629 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
630 } else if (proj_nr == pn_DivMod_res_mod) {
631 const ir_node *a = get_Proj_pred(n);
632 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
635 } /* computed_value_Proj_DivMod */
638 * Return the value of a Proj(Div).
640 static tarval *computed_value_Proj_Div(const ir_node *n)
642 long proj_nr = get_Proj_proj(n);
644 if (proj_nr == pn_Div_res) {
645 const ir_node *a = get_Proj_pred(n);
646 return do_computed_value_Div(get_Div_left(a), get_Div_right(a));
649 } /* computed_value_Proj_Div */
652 * Return the value of a Proj(Mod).
654 static tarval *computed_value_Proj_Mod(const ir_node *n)
656 long proj_nr = get_Proj_proj(n);
658 if (proj_nr == pn_Mod_res) {
659 const ir_node *a = get_Proj_pred(n);
660 return do_computed_value_Mod(get_Mod_left(a), get_Mod_right(a));
663 } /* computed_value_Proj_Mod */
666 * Return the value of a Proj(Quot).
668 static tarval *computed_value_Proj_Quot(const ir_node *n)
670 long proj_nr = get_Proj_proj(n);
672 if (proj_nr == pn_Quot_res) {
673 const ir_node *a = get_Proj_pred(n);
674 return do_computed_value_Quot(get_Quot_left(a), get_Quot_right(a));
677 } /* computed_value_Proj_Quot */
680 * Return the value of a Proj.
682 static tarval *computed_value_Proj(const ir_node *proj)
684 ir_node *n = get_Proj_pred(proj);
686 if (n->op->ops.computed_value_Proj != NULL)
687 return n->op->ops.computed_value_Proj(proj);
689 } /* computed_value_Proj */
692 * If the parameter n can be computed, return its value, else tarval_bad.
693 * Performs constant folding.
695 * @param n The node this should be evaluated
697 tarval *computed_value(const ir_node *n)
699 vrp_attr *vrp = vrp_get_info(n);
700 if (vrp && vrp->valid && tarval_cmp(vrp->bits_set, vrp->bits_not_set) == pn_Cmp_Eq) {
701 return vrp->bits_set;
703 if (n->op->ops.computed_value)
704 return n->op->ops.computed_value(n);
706 } /* computed_value */
709 * Set the default computed_value evaluator in an ir_op_ops.
711 * @param code the opcode for the default operation
712 * @param ops the operations initialized
717 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
721 ops->computed_value = computed_value_##a; \
723 #define CASE_PROJ(a) \
725 ops->computed_value_Proj = computed_value_Proj_##a; \
763 } /* firm_set_default_computed_value */
766 * Returns a equivalent block for another block.
767 * If the block has only one predecessor, this is
768 * the equivalent one. If the only predecessor of a block is
769 * the block itself, this is a dead block.
771 * If both predecessors of a block are the branches of a binary
772 * Cond, the equivalent block is Cond's block.
774 * If all predecessors of a block are bad or lies in a dead
775 * block, the current block is dead as well.
777 * Note, that blocks are NEVER turned into Bad's, instead
778 * the dead_block flag is set. So, never test for is_Bad(block),
779 * always use is_dead_Block(block).
781 static ir_node *equivalent_node_Block(ir_node *n)
787 /* don't optimize dead or labeled blocks */
788 if (is_Block_dead(n) || has_Block_entity(n))
791 n_preds = get_Block_n_cfgpreds(n);
793 /* The Block constructor does not call optimize, but mature_immBlock()
794 calls the optimization. */
795 assert(get_Block_matured(n));
797 irg = get_irn_irg(n);
799 /* Straightening: a single entry Block following a single exit Block
800 can be merged, if it is not the Start block. */
801 /* !!! Beware, all Phi-nodes of n must have been optimized away.
802 This should be true, as the block is matured before optimize is called.
803 But what about Phi-cycles with the Phi0/Id that could not be resolved?
804 Remaining Phi nodes are just Ids. */
806 ir_node *pred = skip_Proj(get_Block_cfgpred(n, 0));
809 ir_node *predblock = get_nodes_block(pred);
810 if (predblock == oldn) {
811 /* Jmp jumps into the block it is in -- deal self cycle. */
812 n = set_Block_dead(n);
813 DBG_OPT_DEAD_BLOCK(oldn, n);
814 } else if (get_opt_control_flow_straightening()) {
816 DBG_OPT_STG(oldn, n);
818 } else if (is_Cond(pred)) {
819 ir_node *predblock = get_nodes_block(pred);
820 if (predblock == oldn) {
821 /* Jmp jumps into the block it is in -- deal self cycle. */
822 n = set_Block_dead(n);
823 DBG_OPT_DEAD_BLOCK(oldn, n);
826 } else if ((n_preds == 2) &&
827 (get_opt_control_flow_weak_simplification())) {
828 /* Test whether Cond jumps twice to this block
829 * The more general case which more than 2 predecessors is handles
830 * in optimize_cf(), we handle only this special case for speed here.
832 ir_node *a = get_Block_cfgpred(n, 0);
833 ir_node *b = get_Block_cfgpred(n, 1);
835 if (is_Proj(a) && is_Proj(b)) {
836 ir_node *cond = get_Proj_pred(a);
838 if (cond == get_Proj_pred(b) && is_Cond(cond) &&
839 get_irn_mode(get_Cond_selector(cond)) == mode_b) {
840 /* Also a single entry Block following a single exit Block. Phis have
841 twice the same operand and will be optimized away. */
842 n = get_nodes_block(cond);
843 DBG_OPT_IFSIM1(oldn, a, b, n);
846 } else if (get_opt_unreachable_code() &&
847 (n != get_irg_start_block(irg)) &&
848 (n != get_irg_end_block(irg))) {
851 /* If all inputs are dead, this block is dead too, except if it is
852 the start or end block. This is one step of unreachable code
854 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
855 ir_node *pred = get_Block_cfgpred(n, i);
858 if (is_Bad(pred)) continue;
859 pred_blk = get_nodes_block(skip_Proj(pred));
861 if (is_Block_dead(pred_blk)) continue;
864 /* really found a living input */
869 n = set_Block_dead(n);
870 DBG_OPT_DEAD_BLOCK(oldn, n);
875 } /* equivalent_node_Block */
878 * Returns a equivalent node for a Jmp, a Bad :-)
879 * Of course this only happens if the Block of the Jmp is dead.
881 static ir_node *equivalent_node_Jmp(ir_node *n)
885 /* unreachable code elimination */
886 if (is_Block_dead(get_nodes_block(n))) {
887 ir_graph *irg = get_irn_irg(n);
888 n = get_irg_bad(irg);
889 DBG_OPT_DEAD_BLOCK(oldn, n);
892 } /* equivalent_node_Jmp */
894 /** Raise is handled in the same way as Jmp. */
895 #define equivalent_node_Raise equivalent_node_Jmp
898 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
899 See transform_node_Proj_Cond(). */
902 * Optimize operations that are commutative and have neutral 0,
903 * so a op 0 = 0 op a = a.
905 static ir_node *equivalent_node_neutral_zero(ir_node *n)
909 ir_node *a = get_binop_left(n);
910 ir_node *b = get_binop_right(n);
915 /* After running compute_node there is only one constant predecessor.
916 Find this predecessors value and remember the other node: */
917 if ((tv = value_of(a)) != tarval_bad) {
919 } else if ((tv = value_of(b)) != tarval_bad) {
924 /* If this predecessors constant value is zero, the operation is
925 * unnecessary. Remove it.
927 * Beware: If n is a Add, the mode of on and n might be different
928 * which happens in this rare construction: NULL + 3.
929 * Then, a Conv would be needed which we cannot include here.
931 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
934 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
938 } /* equivalent_node_neutral_zero */
941 * Eor is commutative and has neutral 0.
943 static ir_node *equivalent_node_Eor(ir_node *n)
949 n = equivalent_node_neutral_zero(n);
950 if (n != oldn) return n;
953 b = get_Eor_right(n);
956 ir_node *aa = get_Eor_left(a);
957 ir_node *ab = get_Eor_right(a);
960 /* (a ^ b) ^ a -> b */
962 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
964 } else if (ab == b) {
965 /* (a ^ b) ^ b -> a */
967 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
972 ir_node *ba = get_Eor_left(b);
973 ir_node *bb = get_Eor_right(b);
976 /* a ^ (a ^ b) -> b */
978 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
980 } else if (bb == a) {
981 /* a ^ (b ^ a) -> b */
983 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
991 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
993 * The second one looks strange, but this construct
994 * is used heavily in the LCC sources :-).
996 * Beware: The Mode of an Add may be different than the mode of its
997 * predecessors, so we could not return a predecessors in all cases.
999 static ir_node *equivalent_node_Add(ir_node *n)
1002 ir_node *left, *right;
1003 ir_mode *mode = get_irn_mode(n);
1005 n = equivalent_node_neutral_zero(n);
1009 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1010 if (mode_is_float(mode)) {
1011 ir_graph *irg = get_irn_irg(n);
1012 if (get_irg_fp_model(irg) & fp_strict_algebraic)
1016 left = get_Add_left(n);
1017 right = get_Add_right(n);
1020 if (get_Sub_right(left) == right) {
1023 n = get_Sub_left(left);
1024 if (mode == get_irn_mode(n)) {
1025 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
1030 if (is_Sub(right)) {
1031 if (get_Sub_right(right) == left) {
1034 n = get_Sub_left(right);
1035 if (mode == get_irn_mode(n)) {
1036 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
1042 } /* equivalent_node_Add */
1045 * optimize operations that are not commutative but have neutral 0 on left,
1048 static ir_node *equivalent_node_left_zero(ir_node *n)
1052 ir_node *a = get_binop_left(n);
1053 ir_node *b = get_binop_right(n);
1054 tarval *tb = value_of(b);
1056 if (tarval_is_null(tb)) {
1059 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1062 } /* equivalent_node_left_zero */
1064 #define equivalent_node_Shl equivalent_node_left_zero
1065 #define equivalent_node_Shr equivalent_node_left_zero
1066 #define equivalent_node_Shrs equivalent_node_left_zero
1067 #define equivalent_node_Rotl equivalent_node_left_zero
1070 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
1072 * The second one looks strange, but this construct
1073 * is used heavily in the LCC sources :-).
1075 * Beware: The Mode of a Sub may be different than the mode of its
1076 * predecessors, so we could not return a predecessors in all cases.
1078 static ir_node *equivalent_node_Sub(ir_node *n)
1082 ir_mode *mode = get_irn_mode(n);
1085 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1086 if (mode_is_float(mode)) {
1087 ir_graph *irg = get_irn_irg(n);
1088 if (get_irg_fp_model(irg) & fp_strict_algebraic)
1092 b = get_Sub_right(n);
1095 /* Beware: modes might be different */
1096 if (tarval_is_null(tb)) {
1097 ir_node *a = get_Sub_left(n);
1098 if (mode == get_irn_mode(a)) {
1101 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1105 } /* equivalent_node_Sub */
1109 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
1112 * -(-a) == a, but might overflow two times.
1113 * We handle it anyway here but the better way would be a
1114 * flag. This would be needed for Pascal for instance.
1116 static ir_node *equivalent_node_idempotent_unop(ir_node *n)
1119 ir_node *pred = get_unop_op(n);
1121 /* optimize symmetric unop */
1122 if (get_irn_op(pred) == get_irn_op(n)) {
1123 n = get_unop_op(pred);
1124 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
1127 } /* equivalent_node_idempotent_unop */
1129 /** Optimize Not(Not(x)) == x. */
1130 #define equivalent_node_Not equivalent_node_idempotent_unop
1132 /** -(-x) == x ??? Is this possible or can --x raise an
1133 out of bounds exception if min =! max? */
1134 #define equivalent_node_Minus equivalent_node_idempotent_unop
1137 * Optimize a * 1 = 1 * a = a.
1139 static ir_node *equivalent_node_Mul(ir_node *n)
1142 ir_node *a = get_Mul_left(n);
1144 /* we can handle here only the n * n = n bit cases */
1145 if (get_irn_mode(n) == get_irn_mode(a)) {
1146 ir_node *b = get_Mul_right(n);
1150 * Mul is commutative and has again an other neutral element.
1151 * Constants are place right, so check this case first.
1154 if (tarval_is_one(tv)) {
1156 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1159 if (tarval_is_one(tv)) {
1161 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1166 } /* equivalent_node_Mul */
1169 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1171 static ir_node *equivalent_node_Or(ir_node *n)
1175 ir_node *a = get_Or_left(n);
1176 ir_node *b = get_Or_right(n);
1180 n = a; /* Or has it's own neutral element */
1181 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1184 /* constants are cormalized to right, check this site first */
1186 if (tarval_is_null(tv)) {
1188 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1192 if (tarval_is_null(tv)) {
1194 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1199 } /* equivalent_node_Or */
1202 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1204 static ir_node *equivalent_node_And(ir_node *n)
1208 ir_node *a = get_And_left(n);
1209 ir_node *b = get_And_right(n);
1213 n = a; /* And has it's own neutral element */
1214 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1217 /* constants are normalized to right, check this site first */
1219 if (tarval_is_all_one(tv)) {
1221 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1224 if (tv != get_tarval_bad()) {
1225 ir_mode *mode = get_irn_mode(n);
1226 if (!mode_is_signed(mode) && is_Conv(a)) {
1227 ir_node *convop = get_Conv_op(a);
1228 ir_mode *convopmode = get_irn_mode(convop);
1229 if (!mode_is_signed(convopmode)) {
1230 if (tarval_is_all_one(tarval_convert_to(tv, convopmode))) {
1231 /* Conv(X) & all_one(mode(X)) = Conv(X) */
1233 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1240 if (tarval_is_all_one(tv)) {
1242 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1246 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1249 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1254 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1257 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1262 } /* equivalent_node_And */
1265 * Try to remove useless Conv's:
1267 static ir_node *equivalent_node_Conv(ir_node *n)
1270 ir_node *a = get_Conv_op(n);
1272 ir_mode *n_mode = get_irn_mode(n);
1273 ir_mode *a_mode = get_irn_mode(a);
1276 if (n_mode == a_mode) { /* No Conv necessary */
1277 if (get_Conv_strict(n)) {
1280 /* neither Minus nor Abs nor Confirm change the precision,
1281 so we can "look-through" */
1284 p = get_Minus_op(p);
1285 } else if (is_Abs(p)) {
1287 } else if (is_Confirm(p)) {
1288 p = get_Confirm_value(p);
1294 if (is_Conv(p) && get_Conv_strict(p)) {
1295 /* we known already, that a_mode == n_mode, and neither
1296 Abs nor Minus change the mode, so the second Conv
1298 assert(get_irn_mode(p) == n_mode);
1300 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1304 ir_node *pred = get_Proj_pred(p);
1305 if (is_Load(pred)) {
1306 /* Loads always return with the exact precision of n_mode */
1307 assert(get_Load_mode(pred) == n_mode);
1309 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1312 if (is_Proj(pred) && get_Proj_proj(pred) == pn_Start_T_args) {
1313 pred = get_Proj_pred(pred);
1314 if (is_Start(pred)) {
1315 /* Arguments always return with the exact precision,
1316 as strictConv's are place before Call -- if the
1317 caller was compiled with the same setting.
1318 Otherwise, the semantics is probably still right. */
1319 assert(get_irn_mode(p) == n_mode);
1321 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1327 /* special case: the immediate predecessor is also a Conv */
1328 if (! get_Conv_strict(a)) {
1329 /* first one is not strict, kick it */
1331 a_mode = get_irn_mode(a);
1335 /* else both are strict conv, second is superfluous */
1337 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1342 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1345 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1346 ir_node *b = get_Conv_op(a);
1347 ir_mode *b_mode = get_irn_mode(b);
1349 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1350 /* both are strict conv */
1351 if (smaller_mode(a_mode, n_mode)) {
1352 /* both are strict, but the first is smaller, so
1353 the second cannot remove more precision, remove the
1355 set_Conv_strict(n, 0);
1358 if (n_mode == b_mode) {
1359 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1360 if (n_mode == mode_b) {
1361 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1362 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1364 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1365 if (values_in_mode(b_mode, a_mode)) {
1366 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1367 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1372 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1373 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1374 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1375 unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode);
1377 if (float_mantissa >= int_mantissa) {
1379 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1384 if (smaller_mode(b_mode, a_mode)) {
1385 if (get_Conv_strict(n))
1386 set_Conv_strict(b, 1);
1387 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1388 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1395 } /* equivalent_node_Conv */
1398 * A Cast may be removed if the type of the previous node
1399 * is already the type of the Cast.
1401 static ir_node *equivalent_node_Cast(ir_node *n)
1404 ir_node *pred = get_Cast_op(n);
1406 if (get_irn_type(pred) == get_Cast_type(n)) {
1408 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1411 } /* equivalent_node_Cast */
1414 * - fold Phi-nodes, iff they have only one predecessor except
1417 static ir_node *equivalent_node_Phi(ir_node *n)
1423 ir_node *first_val = NULL; /* to shutup gcc */
1425 if (!get_opt_normalize()) return n;
1427 n_preds = get_Phi_n_preds(n);
1429 block = get_nodes_block(n);
1431 if (is_Block_dead(block)) {
1432 ir_graph *irg = get_irn_irg(n);
1433 return get_irg_bad(irg);
1436 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1438 /* Find first non-self-referencing input */
1439 for (i = 0; i < n_preds; ++i) {
1440 first_val = get_Phi_pred(n, i);
1441 if ( (first_val != n) /* not self pointer */
1443 /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad
1444 * predecessors. Then, Phi nodes in dead code might be removed, causing
1445 * nodes pointing to themself (Add's for instance).
1446 * This is really bad and causes endless recursions in several
1447 * code pathes, so we do NOT optimize such a code.
1448 * This is not that bad as it sounds, optimize_cf() removes bad control flow
1449 * (and bad Phi predecessors), so live code is optimized later.
1451 && (! is_Bad(get_Block_cfgpred(block, i)))
1453 ) { /* value not dead */
1454 break; /* then found first value. */
1459 ir_graph *irg = get_irn_irg(n);
1460 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1461 return get_irg_bad(irg);
1464 /* search for rest of inputs, determine if any of these
1465 are non-self-referencing */
1466 while (++i < n_preds) {
1467 ir_node *scnd_val = get_Phi_pred(n, i);
1468 if ( (scnd_val != n)
1469 && (scnd_val != first_val)
1472 && (! is_Bad(get_Block_cfgpred(block, i)))
1480 /* Fold, if no multiple distinct non-self-referencing inputs */
1482 DBG_OPT_PHI(oldn, n);
1485 } /* equivalent_node_Phi */
1488 * Several optimizations:
1489 * - fold Sync-nodes, iff they have only one predecessor except
1492 static ir_node *equivalent_node_Sync(ir_node *n)
1494 int arity = get_Sync_n_preds(n);
1497 for (i = 0; i < arity;) {
1498 ir_node *pred = get_Sync_pred(n, i);
1501 /* Remove Bad predecessors */
1508 /* Remove duplicate predecessors */
1514 if (get_Sync_pred(n, j) == pred) {
1523 ir_graph *irg = get_irn_irg(n);
1524 return get_irg_bad(irg);
1526 if (arity == 1) return get_Sync_pred(n, 0);
1528 } /* equivalent_node_Sync */
1531 * Optimize Proj(Tuple).
1533 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj)
1535 ir_node *oldn = proj;
1536 ir_node *tuple = get_Proj_pred(proj);
1538 /* Remove the Tuple/Proj combination. */
1539 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1540 DBG_OPT_TUPLE(oldn, tuple, proj);
1543 } /* equivalent_node_Proj_Tuple */
1546 * Optimize a / 1 = a.
1548 static ir_node *equivalent_node_Proj_Div(ir_node *proj)
1550 ir_node *oldn = proj;
1551 ir_node *div = get_Proj_pred(proj);
1552 ir_node *b = get_Div_right(div);
1553 tarval *tb = value_of(b);
1555 /* Div is not commutative. */
1556 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1557 switch (get_Proj_proj(proj)) {
1559 proj = get_Div_mem(div);
1560 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1564 proj = get_Div_left(div);
1565 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1569 /* we cannot replace the exception Proj's here, this is done in
1570 transform_node_Proj_Div() */
1575 } /* equivalent_node_Proj_Div */
1578 * Optimize a / 1.0 = a.
1580 static ir_node *equivalent_node_Proj_Quot(ir_node *proj)
1582 ir_node *oldn = proj;
1583 ir_node *quot = get_Proj_pred(proj);
1584 ir_node *b = get_Quot_right(quot);
1585 tarval *tb = value_of(b);
1587 /* Div is not commutative. */
1588 if (tarval_is_one(tb)) { /* Quot(x, 1) == x */
1589 switch (get_Proj_proj(proj)) {
1591 proj = get_Quot_mem(quot);
1592 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1596 proj = get_Quot_left(quot);
1597 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1601 /* we cannot replace the exception Proj's here, this is done in
1602 transform_node_Proj_Quot() */
1607 } /* equivalent_node_Proj_Quot */
1610 * Optimize a / 1 = a.
1612 static ir_node *equivalent_node_Proj_DivMod(ir_node *proj)
1614 ir_node *oldn = proj;
1615 ir_node *divmod = get_Proj_pred(proj);
1616 ir_node *b = get_DivMod_right(divmod);
1617 tarval *tb = value_of(b);
1619 /* Div is not commutative. */
1620 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1621 switch (get_Proj_proj(proj)) {
1623 proj = get_DivMod_mem(divmod);
1624 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1627 case pn_DivMod_res_div:
1628 proj = get_DivMod_left(divmod);
1629 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1633 /* we cannot replace the exception Proj's here, this is done in
1634 transform_node_Proj_DivMod().
1635 Note further that the pn_DivMod_res_div case is handled in
1636 computed_value_Proj(). */
1641 } /* equivalent_node_Proj_DivMod */
1644 * Optimize CopyB(mem, x, x) into a Nop.
1646 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj)
1648 ir_node *oldn = proj;
1649 ir_node *copyb = get_Proj_pred(proj);
1650 ir_node *a = get_CopyB_dst(copyb);
1651 ir_node *b = get_CopyB_src(copyb);
1654 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1655 switch (get_Proj_proj(proj)) {
1657 proj = get_CopyB_mem(copyb);
1658 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1661 case pn_CopyB_X_except: {
1662 ir_graph *irg = get_irn_irg(proj);
1663 DBG_OPT_EXC_REM(proj);
1664 proj = get_irg_bad(irg);
1670 } /* equivalent_node_Proj_CopyB */
1673 * Optimize Bounds(idx, idx, upper) into idx.
1675 static ir_node *equivalent_node_Proj_Bound(ir_node *proj)
1677 ir_node *oldn = proj;
1678 ir_node *bound = get_Proj_pred(proj);
1679 ir_node *idx = get_Bound_index(bound);
1680 ir_node *pred = skip_Proj(idx);
1683 if (idx == get_Bound_lower(bound))
1685 else if (is_Bound(pred)) {
1687 * idx was Bounds checked in the same MacroBlock previously,
1688 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1690 ir_node *lower = get_Bound_lower(bound);
1691 ir_node *upper = get_Bound_upper(bound);
1692 if (get_Bound_lower(pred) == lower &&
1693 get_Bound_upper(pred) == upper &&
1694 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
1696 * One could expect that we simply return the previous
1697 * Bound here. However, this would be wrong, as we could
1698 * add an exception Proj to a new location then.
1699 * So, we must turn in into a tuple.
1705 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1706 switch (get_Proj_proj(proj)) {
1708 DBG_OPT_EXC_REM(proj);
1709 proj = get_Bound_mem(bound);
1711 case pn_Bound_X_except: {
1712 ir_graph *irg = get_irn_irg(proj);
1713 DBG_OPT_EXC_REM(proj);
1714 proj = get_irg_bad(irg);
1719 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1722 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1727 } /* equivalent_node_Proj_Bound */
1730 * Optimize an Exception Proj(Load) with a non-null address.
1732 static ir_node *equivalent_node_Proj_Load(ir_node *proj)
1734 if (get_opt_ldst_only_null_ptr_exceptions()) {
1735 if (get_irn_mode(proj) == mode_X) {
1736 ir_node *load = get_Proj_pred(proj);
1738 /* get the Load address */
1739 const ir_node *addr = get_Load_ptr(load);
1740 const ir_node *confirm;
1742 if (value_not_null(addr, &confirm)) {
1743 if (get_Proj_proj(proj) == pn_Load_X_except) {
1744 ir_graph *irg = get_irn_irg(proj);
1745 DBG_OPT_EXC_REM(proj);
1746 return get_irg_bad(irg);
1752 } /* equivalent_node_Proj_Load */
1755 * Optimize an Exception Proj(Store) with a non-null address.
1757 static ir_node *equivalent_node_Proj_Store(ir_node *proj)
1759 if (get_opt_ldst_only_null_ptr_exceptions()) {
1760 if (get_irn_mode(proj) == mode_X) {
1761 ir_node *store = get_Proj_pred(proj);
1763 /* get the load/store address */
1764 const ir_node *addr = get_Store_ptr(store);
1765 const ir_node *confirm;
1767 if (value_not_null(addr, &confirm)) {
1768 if (get_Proj_proj(proj) == pn_Store_X_except) {
1769 ir_graph *irg = get_irn_irg(proj);
1770 DBG_OPT_EXC_REM(proj);
1771 return get_irg_bad(irg);
1777 } /* equivalent_node_Proj_Store */
1780 * Does all optimizations on nodes that must be done on it's Proj's
1781 * because of creating new nodes.
1783 static ir_node *equivalent_node_Proj(ir_node *proj)
1785 ir_node *n = get_Proj_pred(proj);
1787 if (get_irn_mode(proj) == mode_X) {
1788 if (is_Block_dead(get_nodes_block(n))) {
1789 /* Remove dead control flow -- early gigo(). */
1790 ir_graph *irg = get_irn_irg(proj);
1791 return get_irg_bad(irg);
1794 if (n->op->ops.equivalent_node_Proj)
1795 return n->op->ops.equivalent_node_Proj(proj);
1797 } /* equivalent_node_Proj */
1802 static ir_node *equivalent_node_Id(ir_node *n)
1810 DBG_OPT_ID(oldn, n);
1812 } /* equivalent_node_Id */
1817 static ir_node *equivalent_node_Mux(ir_node *n)
1819 ir_node *oldn = n, *sel = get_Mux_sel(n);
1821 tarval *ts = value_of(sel);
1823 /* Mux(true, f, t) == t */
1824 if (ts == tarval_b_true) {
1825 n = get_Mux_true(n);
1826 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1829 /* Mux(false, f, t) == f */
1830 if (ts == tarval_b_false) {
1831 n = get_Mux_false(n);
1832 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1835 n_t = get_Mux_true(n);
1836 n_f = get_Mux_false(n);
1838 /* Mux(v, x, T) == x */
1839 if (is_Unknown(n_f)) {
1841 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1844 /* Mux(v, T, x) == x */
1845 if (is_Unknown(n_t)) {
1847 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1851 /* Mux(v, x, x) == x */
1854 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1857 if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1858 ir_node *cmp = get_Proj_pred(sel);
1859 long proj_nr = get_Proj_proj(sel);
1860 ir_node *f = get_Mux_false(n);
1861 ir_node *t = get_Mux_true(n);
1864 * Note further that these optimization work even for floating point
1865 * with NaN's because -NaN == NaN.
1866 * However, if +0 and -0 is handled differently, we cannot use the first one.
1869 ir_node *const cmp_l = get_Cmp_left(cmp);
1870 ir_node *const cmp_r = get_Cmp_right(cmp);
1874 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1875 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1877 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1884 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1885 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1887 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1894 * Note: normalization puts the constant on the right side,
1895 * so we check only one case.
1897 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1898 /* Mux(t CMP 0, X, t) */
1899 if (is_Minus(f) && get_Minus_op(f) == t) {
1900 /* Mux(t CMP 0, -t, t) */
1901 if (proj_nr == pn_Cmp_Eq) {
1902 /* Mux(t == 0, -t, t) ==> -t */
1904 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1905 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1906 /* Mux(t != 0, -t, t) ==> t */
1908 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1915 } /* equivalent_node_Mux */
1918 * Remove Confirm nodes if setting is on.
1919 * Replace Confirms(x, '=', Constlike) by Constlike.
1921 static ir_node *equivalent_node_Confirm(ir_node *n)
1923 ir_node *pred = get_Confirm_value(n);
1924 pn_Cmp pnc = get_Confirm_cmp(n);
1926 while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1928 * rare case: two identical Confirms one after another,
1929 * replace the second one with the first.
1932 pred = get_Confirm_value(n);
1933 pnc = get_Confirm_cmp(n);
1935 if (get_opt_remove_confirm())
1936 return get_Confirm_value(n);
1941 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1942 * perform no actual computation, as, e.g., the Id nodes. It does not create
1943 * new nodes. It is therefore safe to free n if the node returned is not n.
1944 * If a node returns a Tuple we can not just skip it. If the size of the
1945 * in array fits, we transform n into a tuple (e.g., Div).
1947 ir_node *equivalent_node(ir_node *n)
1949 if (n->op->ops.equivalent_node)
1950 return n->op->ops.equivalent_node(n);
1952 } /* equivalent_node */
1955 * Sets the default equivalent node operation for an ir_op_ops.
1957 * @param code the opcode for the default operation
1958 * @param ops the operations initialized
1963 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1967 ops->equivalent_node = equivalent_node_##a; \
1969 #define CASE_PROJ(a) \
1971 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
2014 } /* firm_set_default_equivalent_node */
2017 * Returns non-zero if a node is a Phi node
2018 * with all predecessors constant.
2020 static int is_const_Phi(ir_node *n)
2024 if (! is_Phi(n) || get_irn_arity(n) == 0)
2026 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
2027 if (! is_Const(get_irn_n(n, i)))
2031 } /* is_const_Phi */
2033 typedef tarval *(*tarval_sub_type)(tarval *a, tarval *b, ir_mode *mode);
2034 typedef tarval *(*tarval_binop_type)(tarval *a, tarval *b);
2037 * in reality eval_func should be tarval (*eval_func)() but incomplete
2038 * declarations are bad style and generate noisy warnings
2040 typedef void (*eval_func)(void);
2043 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
2045 static tarval *do_eval(eval_func eval, tarval *a, tarval *b, ir_mode *mode)
2047 if (eval == (eval_func) tarval_sub) {
2048 tarval_sub_type func = (tarval_sub_type)eval;
2050 return func(a, b, mode);
2052 tarval_binop_type func = (tarval_binop_type)eval;
2059 * Apply an evaluator on a binop with a constant operators (and one Phi).
2061 * @param phi the Phi node
2062 * @param other the other operand
2063 * @param eval an evaluator function
2064 * @param mode the mode of the result, may be different from the mode of the Phi!
2065 * @param left if non-zero, other is the left operand, else the right
2067 * @return a new Phi node if the conversion was successful, NULL else
2069 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, eval_func eval, ir_mode *mode, int left)
2075 int i, n = get_irn_arity(phi);
2077 NEW_ARR_A(void *, res, n);
2079 for (i = 0; i < n; ++i) {
2080 pred = get_irn_n(phi, i);
2081 tv = get_Const_tarval(pred);
2082 tv = do_eval(eval, other, tv, mode);
2084 if (tv == tarval_bad) {
2085 /* folding failed, bad */
2091 for (i = 0; i < n; ++i) {
2092 pred = get_irn_n(phi, i);
2093 tv = get_Const_tarval(pred);
2094 tv = do_eval(eval, tv, other, mode);
2096 if (tv == tarval_bad) {
2097 /* folding failed, bad */
2103 irg = get_irn_irg(phi);
2104 for (i = 0; i < n; ++i) {
2105 pred = get_irn_n(phi, i);
2106 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2108 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2109 } /* apply_binop_on_phi */
2112 * Apply an evaluator on a binop with two constant Phi.
2114 * @param a the left Phi node
2115 * @param b the right Phi node
2116 * @param eval an evaluator function
2117 * @param mode the mode of the result, may be different from the mode of the Phi!
2119 * @return a new Phi node if the conversion was successful, NULL else
2121 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, eval_func eval, ir_mode *mode)
2123 tarval *tv_l, *tv_r, *tv;
2129 if (get_nodes_block(a) != get_nodes_block(b))
2132 n = get_irn_arity(a);
2133 NEW_ARR_A(void *, res, n);
2135 for (i = 0; i < n; ++i) {
2136 pred = get_irn_n(a, i);
2137 tv_l = get_Const_tarval(pred);
2138 pred = get_irn_n(b, i);
2139 tv_r = get_Const_tarval(pred);
2140 tv = do_eval(eval, tv_l, tv_r, mode);
2142 if (tv == tarval_bad) {
2143 /* folding failed, bad */
2148 irg = get_irn_irg(a);
2149 for (i = 0; i < n; ++i) {
2150 pred = get_irn_n(a, i);
2151 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2153 return new_r_Phi(get_nodes_block(a), n, (ir_node **)res, mode);
2154 } /* apply_binop_on_2_phis */
2157 * Apply an evaluator on a unop with a constant operator (a Phi).
2159 * @param phi the Phi node
2160 * @param eval an evaluator function
2162 * @return a new Phi node if the conversion was successful, NULL else
2164 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *))
2171 int i, n = get_irn_arity(phi);
2173 NEW_ARR_A(void *, res, n);
2174 for (i = 0; i < n; ++i) {
2175 pred = get_irn_n(phi, i);
2176 tv = get_Const_tarval(pred);
2179 if (tv == tarval_bad) {
2180 /* folding failed, bad */
2185 mode = get_irn_mode(phi);
2186 irg = get_irn_irg(phi);
2187 for (i = 0; i < n; ++i) {
2188 pred = get_irn_n(phi, i);
2189 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2191 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2192 } /* apply_unop_on_phi */
2195 * Apply a conversion on a constant operator (a Phi).
2197 * @param phi the Phi node
2199 * @return a new Phi node if the conversion was successful, NULL else
2201 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode)
2207 int i, n = get_irn_arity(phi);
2209 NEW_ARR_A(void *, res, n);
2210 for (i = 0; i < n; ++i) {
2211 pred = get_irn_n(phi, i);
2212 tv = get_Const_tarval(pred);
2213 tv = tarval_convert_to(tv, mode);
2215 if (tv == tarval_bad) {
2216 /* folding failed, bad */
2221 irg = get_irn_irg(phi);
2222 for (i = 0; i < n; ++i) {
2223 pred = get_irn_n(phi, i);
2224 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2226 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2227 } /* apply_conv_on_phi */
2230 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
2231 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
2232 * If possible, remove the Conv's.
2234 static ir_node *transform_node_AddSub(ir_node *n)
2236 ir_mode *mode = get_irn_mode(n);
2238 if (mode_is_reference(mode)) {
2239 ir_node *left = get_binop_left(n);
2240 ir_node *right = get_binop_right(n);
2241 unsigned ref_bits = get_mode_size_bits(mode);
2243 if (is_Conv(left)) {
2244 ir_mode *lmode = get_irn_mode(left);
2245 unsigned bits = get_mode_size_bits(lmode);
2247 if (ref_bits == bits &&
2248 mode_is_int(lmode) &&
2249 get_mode_arithmetic(lmode) == irma_twos_complement) {
2250 ir_node *pre = get_Conv_op(left);
2251 ir_mode *pre_mode = get_irn_mode(pre);
2253 if (mode_is_int(pre_mode) &&
2254 get_mode_size_bits(pre_mode) == bits &&
2255 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2256 /* ok, this conv just changes to sign, moreover the calculation
2257 * is done with same number of bits as our address mode, so
2258 * we can ignore the conv as address calculation can be viewed
2259 * as either signed or unsigned
2261 set_binop_left(n, pre);
2266 if (is_Conv(right)) {
2267 ir_mode *rmode = get_irn_mode(right);
2268 unsigned bits = get_mode_size_bits(rmode);
2270 if (ref_bits == bits &&
2271 mode_is_int(rmode) &&
2272 get_mode_arithmetic(rmode) == irma_twos_complement) {
2273 ir_node *pre = get_Conv_op(right);
2274 ir_mode *pre_mode = get_irn_mode(pre);
2276 if (mode_is_int(pre_mode) &&
2277 get_mode_size_bits(pre_mode) == bits &&
2278 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2279 /* ok, this conv just changes to sign, moreover the calculation
2280 * is done with same number of bits as our address mode, so
2281 * we can ignore the conv as address calculation can be viewed
2282 * as either signed or unsigned
2284 set_binop_right(n, pre);
2289 /* let address arithmetic use unsigned modes */
2290 if (is_Const(right)) {
2291 ir_mode *rmode = get_irn_mode(right);
2293 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2294 /* convert a AddP(P, *s) into AddP(P, *u) */
2295 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2297 ir_node *pre = new_r_Conv(get_nodes_block(n), right, nm);
2298 set_binop_right(n, pre);
2304 } /* transform_node_AddSub */
2306 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2309 if (is_Const(b) && is_const_Phi(a)) { \
2310 /* check for Op(Phi, Const) */ \
2311 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2313 else if (is_Const(a) && is_const_Phi(b)) { \
2314 /* check for Op(Const, Phi) */ \
2315 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2317 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2318 /* check for Op(Phi, Phi) */ \
2319 c = apply_binop_on_2_phis(a, b, eval, mode); \
2322 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2327 #define HANDLE_UNOP_PHI(eval, a, c) \
2330 if (is_const_Phi(a)) { \
2331 /* check for Op(Phi) */ \
2332 c = apply_unop_on_phi(a, eval); \
2334 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2341 * Do the AddSub optimization, then Transform
2342 * Constant folding on Phi
2343 * Add(a,a) -> Mul(a, 2)
2344 * Add(Mul(a, x), a) -> Mul(a, x+1)
2345 * if the mode is integer or float.
2346 * Transform Add(a,-b) into Sub(a,b).
2347 * Reassociation might fold this further.
2349 static ir_node *transform_node_Add(ir_node *n)
2352 ir_node *a, *b, *c, *oldn = n;
2353 vrp_attr *a_vrp, *b_vrp;
2355 n = transform_node_AddSub(n);
2357 a = get_Add_left(n);
2358 b = get_Add_right(n);
2360 mode = get_irn_mode(n);
2362 if (mode_is_reference(mode)) {
2363 ir_mode *lmode = get_irn_mode(a);
2365 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2366 /* an Add(a, NULL) is a hidden Conv */
2367 dbg_info *dbg = get_irn_dbg_info(n);
2368 return new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2372 HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode);
2374 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2375 if (mode_is_float(mode)) {
2376 ir_graph *irg = get_irn_irg(n);
2377 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2381 if (mode_is_num(mode)) {
2382 ir_graph *irg = get_irn_irg(n);
2383 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2384 if (!is_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP)
2385 && a == b && mode_is_int(mode)) {
2386 ir_node *block = get_nodes_block(n);
2389 get_irn_dbg_info(n),
2392 new_Const_long(mode, 2),
2394 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2399 get_irn_dbg_info(n),
2404 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2409 get_irn_dbg_info(n),
2414 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2417 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2418 /* Here we rely on constants be on the RIGHT side */
2420 ir_node *op = get_Not_op(a);
2422 if (is_Const(b) && is_Const_one(b)) {
2424 ir_node *blk = get_nodes_block(n);
2425 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, mode);
2426 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2431 n = new_Const(get_mode_minus_one(mode));
2432 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2437 ir_node *op = get_Not_op(b);
2441 n = new_Const(get_mode_minus_one(mode));
2442 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2449 a_vrp = vrp_get_info(a);
2450 b_vrp = vrp_get_info(b);
2452 if (a_vrp && b_vrp) {
2453 tarval *c = tarval_and(
2454 a_vrp->bits_not_set,
2458 if (tarval_is_null(c)) {
2459 dbg_info *dbgi = get_irn_dbg_info(n);
2460 return new_rd_Or(dbgi, get_nodes_block(n),
2465 } /* transform_node_Add */
2468 * returns -cnst or NULL if impossible
2470 static ir_node *const_negate(ir_node *cnst)
2472 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2473 dbg_info *dbgi = get_irn_dbg_info(cnst);
2474 ir_graph *irg = get_irn_irg(cnst);
2475 if (tv == tarval_bad) return NULL;
2476 return new_rd_Const(dbgi, irg, tv);
2480 * Do the AddSub optimization, then Transform
2481 * Constant folding on Phi
2482 * Sub(0,a) -> Minus(a)
2483 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2484 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2485 * Sub(Add(a, x), x) -> a
2486 * Sub(x, Add(x, a)) -> -a
2487 * Sub(x, Const) -> Add(x, -Const)
2489 static ir_node *transform_node_Sub(ir_node *n)
2495 n = transform_node_AddSub(n);
2497 a = get_Sub_left(n);
2498 b = get_Sub_right(n);
2500 mode = get_irn_mode(n);
2502 if (mode_is_int(mode)) {
2503 ir_mode *lmode = get_irn_mode(a);
2505 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2506 /* a Sub(a, NULL) is a hidden Conv */
2507 dbg_info *dbg = get_irn_dbg_info(n);
2508 n = new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2509 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2513 if (mode == lmode &&
2514 get_mode_arithmetic(mode) == irma_twos_complement &&
2516 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2518 dbg_info *dbg = get_irn_dbg_info(n);
2519 n = new_rd_Not(dbg, get_nodes_block(n), b, mode);
2520 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2526 HANDLE_BINOP_PHI((eval_func) tarval_sub, a, b, c, mode);
2528 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2529 if (mode_is_float(mode)) {
2530 ir_graph *irg = get_irn_irg(n);
2531 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2535 if (is_Const(b) && !mode_is_reference(get_irn_mode(b))) {
2536 /* a - C -> a + (-C) */
2537 ir_node *cnst = const_negate(b);
2539 ir_node *block = get_nodes_block(n);
2540 dbg_info *dbgi = get_irn_dbg_info(n);
2542 n = new_rd_Add(dbgi, block, a, cnst, mode);
2543 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2548 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2549 dbg_info *dbg = get_irn_dbg_info(n);
2550 ir_node *block = get_nodes_block(n);
2551 ir_node *left = get_Minus_op(a);
2552 ir_node *add = new_rd_Add(dbg, block, left, b, mode);
2554 n = new_rd_Minus(dbg, block, add, mode);
2555 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2557 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2558 dbg_info *dbg = get_irn_dbg_info(n);
2559 ir_node *block = get_nodes_block(n);
2560 ir_node *right = get_Minus_op(b);
2562 n = new_rd_Add(dbg, block, a, right, mode);
2563 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2565 } else if (is_Sub(b)) {
2566 /* a - (b - c) -> a + (c - b)
2567 * -> (a - b) + c iff (b - c) is a pointer */
2568 dbg_info *s_dbg = get_irn_dbg_info(b);
2569 ir_node *s_block = get_nodes_block(b);
2570 ir_node *s_left = get_Sub_left(b);
2571 ir_node *s_right = get_Sub_right(b);
2572 ir_mode *s_mode = get_irn_mode(b);
2573 if (mode_is_reference(s_mode)) {
2574 ir_node *sub = new_rd_Sub(s_dbg, s_block, a, s_left, mode);
2575 dbg_info *a_dbg = get_irn_dbg_info(n);
2576 ir_node *a_block = get_nodes_block(n);
2579 s_right = new_r_Conv(a_block, s_right, mode);
2580 n = new_rd_Add(a_dbg, a_block, sub, s_right, mode);
2582 ir_node *sub = new_rd_Sub(s_dbg, s_block, s_right, s_left, s_mode);
2583 dbg_info *a_dbg = get_irn_dbg_info(n);
2584 ir_node *a_block = get_nodes_block(n);
2586 n = new_rd_Add(a_dbg, a_block, a, sub, mode);
2588 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2590 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2591 ir_node *m_right = get_Mul_right(b);
2592 if (is_Const(m_right)) {
2593 ir_node *cnst2 = const_negate(m_right);
2594 if (cnst2 != NULL) {
2595 dbg_info *m_dbg = get_irn_dbg_info(b);
2596 ir_node *m_block = get_nodes_block(b);
2597 ir_node *m_left = get_Mul_left(b);
2598 ir_mode *m_mode = get_irn_mode(b);
2599 ir_node *mul = new_rd_Mul(m_dbg, m_block, m_left, cnst2, m_mode);
2600 dbg_info *a_dbg = get_irn_dbg_info(n);
2601 ir_node *a_block = get_nodes_block(n);
2603 n = new_rd_Add(a_dbg, a_block, a, mul, mode);
2604 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2610 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2611 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2613 get_irn_dbg_info(n),
2617 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2621 if (mode_wrap_around(mode)) {
2622 ir_node *left = get_Add_left(a);
2623 ir_node *right = get_Add_right(a);
2625 /* FIXME: Does the Conv's work only for two complement or generally? */
2627 if (mode != get_irn_mode(right)) {
2628 /* This Sub is an effective Cast */
2629 right = new_r_Conv(get_nodes_block(n), right, mode);
2632 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2634 } else if (right == b) {
2635 if (mode != get_irn_mode(left)) {
2636 /* This Sub is an effective Cast */
2637 left = new_r_Conv(get_nodes_block(n), left, mode);
2640 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2646 if (mode_wrap_around(mode)) {
2647 ir_node *left = get_Add_left(b);
2648 ir_node *right = get_Add_right(b);
2650 /* FIXME: Does the Conv's work only for two complement or generally? */
2652 ir_mode *r_mode = get_irn_mode(right);
2654 n = new_r_Minus(get_nodes_block(n), right, r_mode);
2655 if (mode != r_mode) {
2656 /* This Sub is an effective Cast */
2657 n = new_r_Conv(get_nodes_block(n), n, mode);
2659 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2661 } else if (right == a) {
2662 ir_mode *l_mode = get_irn_mode(left);
2664 n = new_r_Minus(get_nodes_block(n), left, l_mode);
2665 if (mode != l_mode) {
2666 /* This Sub is an effective Cast */
2667 n = new_r_Conv(get_nodes_block(n), n, mode);
2669 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2674 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2675 ir_mode *mode = get_irn_mode(a);
2677 if (mode == get_irn_mode(b)) {
2679 ir_node *op_a = get_Conv_op(a);
2680 ir_node *op_b = get_Conv_op(b);
2682 /* check if it's allowed to skip the conv */
2683 ma = get_irn_mode(op_a);
2684 mb = get_irn_mode(op_b);
2686 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2687 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2690 set_Sub_right(n, b);
2696 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2697 if (!is_reassoc_running() && is_Mul(a)) {
2698 ir_node *ma = get_Mul_left(a);
2699 ir_node *mb = get_Mul_right(a);
2702 ir_node *blk = get_nodes_block(n);
2704 get_irn_dbg_info(n),
2708 get_irn_dbg_info(n),
2711 new_Const(get_mode_one(mode)),
2714 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2716 } else if (mb == b) {
2717 ir_node *blk = get_nodes_block(n);
2719 get_irn_dbg_info(n),
2723 get_irn_dbg_info(n),
2726 new_Const(get_mode_one(mode)),
2729 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2733 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2734 ir_node *x = get_Sub_left(a);
2735 ir_node *y = get_Sub_right(a);
2736 ir_node *blk = get_nodes_block(n);
2737 ir_mode *m_b = get_irn_mode(b);
2738 ir_mode *m_y = get_irn_mode(y);
2742 /* Determine the right mode for the Add. */
2745 else if (mode_is_reference(m_b))
2747 else if (mode_is_reference(m_y))
2751 * Both modes are different but none is reference,
2752 * happens for instance in SubP(SubP(P, Iu), Is).
2753 * We have two possibilities here: Cast or ignore.
2754 * Currently we ignore this case.
2759 add = new_r_Add(blk, y, b, add_mode);
2761 n = new_rd_Sub(get_irn_dbg_info(n), blk, x, add, mode);
2762 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2766 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2767 if (is_Const(a) && is_Not(b)) {
2768 /* c - ~X = X + (c+1) */
2769 tarval *tv = get_Const_tarval(a);
2771 tv = tarval_add(tv, get_mode_one(mode));
2772 if (tv != tarval_bad) {
2773 ir_node *blk = get_nodes_block(n);
2774 ir_node *c = new_Const(tv);
2775 n = new_rd_Add(get_irn_dbg_info(n), blk, get_Not_op(b), c, mode);
2776 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2782 } /* transform_node_Sub */
2785 * Several transformation done on n*n=2n bits mul.
2786 * These transformations must be done here because new nodes may be produced.
2788 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode)
2791 ir_node *a = get_Mul_left(n);
2792 ir_node *b = get_Mul_right(n);
2793 tarval *ta = value_of(a);
2794 tarval *tb = value_of(b);
2795 ir_mode *smode = get_irn_mode(a);
2797 if (ta == get_mode_one(smode)) {
2798 /* (L)1 * (L)b = (L)b */
2799 ir_node *blk = get_nodes_block(n);
2800 n = new_rd_Conv(get_irn_dbg_info(n), blk, b, mode);
2801 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2804 else if (ta == get_mode_minus_one(smode)) {
2805 /* (L)-1 * (L)b = (L)b */
2806 ir_node *blk = get_nodes_block(n);
2807 n = new_rd_Minus(get_irn_dbg_info(n), blk, b, smode);
2808 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2809 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2812 if (tb == get_mode_one(smode)) {
2813 /* (L)a * (L)1 = (L)a */
2814 ir_node *blk = get_irn_n(a, -1);
2815 n = new_rd_Conv(get_irn_dbg_info(n), blk, a, mode);
2816 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2819 else if (tb == get_mode_minus_one(smode)) {
2820 /* (L)a * (L)-1 = (L)-a */
2821 ir_node *blk = get_nodes_block(n);
2822 n = new_rd_Minus(get_irn_dbg_info(n), blk, a, smode);
2823 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2824 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2831 * Transform Mul(a,-1) into -a.
2832 * Do constant evaluation of Phi nodes.
2833 * Do architecture dependent optimizations on Mul nodes
2835 static ir_node *transform_node_Mul(ir_node *n)
2837 ir_node *c, *oldn = n;
2838 ir_mode *mode = get_irn_mode(n);
2839 ir_node *a = get_Mul_left(n);
2840 ir_node *b = get_Mul_right(n);
2842 if (is_Bad(a) || is_Bad(b))
2845 if (mode != get_irn_mode(a))
2846 return transform_node_Mul2n(n, mode);
2848 HANDLE_BINOP_PHI((eval_func) tarval_mul, a, b, c, mode);
2850 if (mode_is_signed(mode)) {
2853 if (value_of(a) == get_mode_minus_one(mode))
2855 else if (value_of(b) == get_mode_minus_one(mode))
2858 n = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), r, mode);
2859 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2864 if (is_Const(b)) { /* (-a) * const -> a * -const */
2865 ir_node *cnst = const_negate(b);
2867 dbg_info *dbgi = get_irn_dbg_info(n);
2868 ir_node *block = get_nodes_block(n);
2869 n = new_rd_Mul(dbgi, block, get_Minus_op(a), cnst, mode);
2870 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2873 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2874 dbg_info *dbgi = get_irn_dbg_info(n);
2875 ir_node *block = get_nodes_block(n);
2876 n = new_rd_Mul(dbgi, block, get_Minus_op(a), get_Minus_op(b), mode);
2877 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2879 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2880 ir_node *sub_l = get_Sub_left(b);
2881 ir_node *sub_r = get_Sub_right(b);
2882 dbg_info *dbgi = get_irn_dbg_info(n);
2883 ir_node *block = get_nodes_block(n);
2884 ir_node *new_b = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2885 n = new_rd_Mul(dbgi, block, get_Minus_op(a), new_b, mode);
2886 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2889 } else if (is_Minus(b)) {
2890 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2891 ir_node *sub_l = get_Sub_left(a);
2892 ir_node *sub_r = get_Sub_right(a);
2893 dbg_info *dbgi = get_irn_dbg_info(n);
2894 ir_node *block = get_nodes_block(n);
2895 ir_node *new_a = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2896 n = new_rd_Mul(dbgi, block, new_a, get_Minus_op(b), mode);
2897 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2900 } else if (is_Shl(a)) {
2901 ir_node *const shl_l = get_Shl_left(a);
2902 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2903 /* (1 << x) * b -> b << x */
2904 dbg_info *const dbgi = get_irn_dbg_info(n);
2905 ir_node *const block = get_nodes_block(n);
2906 ir_node *const shl_r = get_Shl_right(a);
2907 n = new_rd_Shl(dbgi, block, b, shl_r, mode);
2908 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2911 } else if (is_Shl(b)) {
2912 ir_node *const shl_l = get_Shl_left(b);
2913 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2914 /* a * (1 << x) -> a << x */
2915 dbg_info *const dbgi = get_irn_dbg_info(n);
2916 ir_node *const block = get_nodes_block(n);
2917 ir_node *const shl_r = get_Shl_right(b);
2918 n = new_rd_Shl(dbgi, block, a, shl_r, mode);
2919 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2923 if (get_mode_arithmetic(mode) == irma_ieee754) {
2925 tarval *tv = get_Const_tarval(a);
2926 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2927 && !tarval_is_negative(tv)) {
2928 /* 2.0 * b = b + b */
2929 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), b, b, mode);
2930 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2934 else if (is_Const(b)) {
2935 tarval *tv = get_Const_tarval(b);
2936 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2937 && !tarval_is_negative(tv)) {
2938 /* a * 2.0 = a + a */
2939 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), a, a, mode);
2940 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2945 return arch_dep_replace_mul_with_shifts(n);
2946 } /* transform_node_Mul */
2949 * Transform a Div Node.
2951 static ir_node *transform_node_Div(ir_node *n)
2953 ir_mode *mode = get_Div_resmode(n);
2954 ir_node *a = get_Div_left(n);
2955 ir_node *b = get_Div_right(n);
2957 const ir_node *dummy;
2959 if (is_Const(b) && is_const_Phi(a)) {
2960 /* check for Div(Phi, Const) */
2961 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0);
2963 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2967 else if (is_Const(a) && is_const_Phi(b)) {
2968 /* check for Div(Const, Phi) */
2969 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1);
2971 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2975 else if (is_const_Phi(a) && is_const_Phi(b)) {
2976 /* check for Div(Phi, Phi) */
2977 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode);
2979 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2986 if (a == b && value_not_zero(a, &dummy)) {
2987 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2988 value = new_Const(get_mode_one(mode));
2989 DBG_OPT_CSTEVAL(n, value);
2992 if (mode_is_signed(mode) && is_Const(b)) {
2993 tarval *tv = get_Const_tarval(b);
2995 if (tv == get_mode_minus_one(mode)) {
2997 value = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
2998 DBG_OPT_CSTEVAL(n, value);
3002 /* Try architecture dependent optimization */
3003 value = arch_dep_replace_div_by_const(n);
3010 /* Turn Div into a tuple (mem, jmp, bad, value) */
3011 mem = get_Div_mem(n);
3012 blk = get_nodes_block(n);
3014 /* skip a potential Pin */
3015 mem = skip_Pin(mem);
3016 turn_into_tuple(n, pn_Div_max);
3017 set_Tuple_pred(n, pn_Div_M, mem);
3018 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(blk));
3019 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
3020 set_Tuple_pred(n, pn_Div_res, value);
3023 } /* transform_node_Div */
3026 * Transform a Mod node.
3028 static ir_node *transform_node_Mod(ir_node *n)
3030 ir_mode *mode = get_Mod_resmode(n);
3031 ir_node *a = get_Mod_left(n);
3032 ir_node *b = get_Mod_right(n);
3036 if (is_Const(b) && is_const_Phi(a)) {
3037 /* check for Div(Phi, Const) */
3038 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0);
3040 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
3044 else if (is_Const(a) && is_const_Phi(b)) {
3045 /* check for Div(Const, Phi) */
3046 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1);
3048 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
3052 else if (is_const_Phi(a) && is_const_Phi(b)) {
3053 /* check for Div(Phi, Phi) */
3054 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode);
3056 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
3063 if (tv != tarval_bad) {
3064 value = new_Const(tv);
3066 DBG_OPT_CSTEVAL(n, value);
3069 ir_node *a = get_Mod_left(n);
3070 ir_node *b = get_Mod_right(n);
3071 const ir_node *dummy;
3073 if (a == b && value_not_zero(a, &dummy)) {
3074 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
3075 value = new_Const(get_mode_null(mode));
3076 DBG_OPT_CSTEVAL(n, value);
3079 if (mode_is_signed(mode) && is_Const(b)) {
3080 tarval *tv = get_Const_tarval(b);
3082 if (tv == get_mode_minus_one(mode)) {
3084 value = new_Const(get_mode_null(mode));
3085 DBG_OPT_CSTEVAL(n, value);
3089 /* Try architecture dependent optimization */
3090 value = arch_dep_replace_mod_by_const(n);
3098 /* Turn Mod into a tuple (mem, jmp, bad, value) */
3099 mem = get_Mod_mem(n);
3100 blk = get_nodes_block(n);
3102 /* skip a potential Pin */
3103 mem = skip_Pin(mem);
3104 turn_into_tuple(n, pn_Mod_max);
3105 set_Tuple_pred(n, pn_Mod_M, mem);
3106 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(blk));
3107 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
3108 set_Tuple_pred(n, pn_Mod_res, value);
3111 } /* transform_node_Mod */
3114 * Transform a DivMod node.
3116 static ir_node *transform_node_DivMod(ir_node *n)
3118 const ir_node *dummy;
3119 ir_node *a = get_DivMod_left(n);
3120 ir_node *b = get_DivMod_right(n);
3121 ir_mode *mode = get_DivMod_resmode(n);
3126 if (is_Const(b) && is_const_Phi(a)) {
3127 /* check for Div(Phi, Const) */
3128 va = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0);
3129 vb = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0);
3131 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
3132 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
3136 else if (is_Const(a) && is_const_Phi(b)) {
3137 /* check for Div(Const, Phi) */
3138 va = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1);
3139 vb = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1);
3141 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
3142 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
3146 else if (is_const_Phi(a) && is_const_Phi(b)) {
3147 /* check for Div(Phi, Phi) */
3148 va = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode);
3149 vb = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode);
3151 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
3152 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
3159 if (tb != tarval_bad) {
3160 if (tb == get_mode_one(get_tarval_mode(tb))) {
3162 vb = new_Const(get_mode_null(mode));
3163 DBG_OPT_CSTEVAL(n, vb);
3165 } else if (ta != tarval_bad) {
3166 tarval *resa, *resb;
3167 resa = tarval_div(ta, tb);
3168 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
3169 Jmp for X result!? */
3170 resb = tarval_mod(ta, tb);
3171 if (resb == tarval_bad) return n; /* Causes exception! */
3172 va = new_Const(resa);
3173 vb = new_Const(resb);
3174 DBG_OPT_CSTEVAL(n, va);
3175 DBG_OPT_CSTEVAL(n, vb);
3177 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
3178 va = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
3179 vb = new_Const(get_mode_null(mode));
3180 DBG_OPT_CSTEVAL(n, va);
3181 DBG_OPT_CSTEVAL(n, vb);
3183 } else { /* Try architecture dependent optimization */
3186 arch_dep_replace_divmod_by_const(&va, &vb, n);
3187 evaluated = va != NULL;
3189 } else if (a == b) {
3190 if (value_not_zero(a, &dummy)) {
3192 va = new_Const(get_mode_one(mode));
3193 vb = new_Const(get_mode_null(mode));
3194 DBG_OPT_CSTEVAL(n, va);
3195 DBG_OPT_CSTEVAL(n, vb);
3198 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
3201 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
3202 /* 0 / non-Const = 0 */
3207 if (evaluated) { /* replace by tuple */
3211 mem = get_DivMod_mem(n);
3212 /* skip a potential Pin */
3213 mem = skip_Pin(mem);
3215 blk = get_nodes_block(n);
3216 turn_into_tuple(n, pn_DivMod_max);
3217 set_Tuple_pred(n, pn_DivMod_M, mem);
3218 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(blk));
3219 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
3220 set_Tuple_pred(n, pn_DivMod_res_div, va);
3221 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
3225 } /* transform_node_DivMod */
3228 * Optimize x / c to x * (1/c)
3230 static ir_node *transform_node_Quot(ir_node *n)
3232 ir_mode *mode = get_Quot_resmode(n);
3235 if (get_mode_arithmetic(mode) == irma_ieee754) {
3236 ir_node *b = get_Quot_right(n);
3237 tarval *tv = value_of(b);
3239 if (tv != tarval_bad) {
3240 int rem = tarval_fp_ops_enabled();
3243 * Floating point constant folding might be disabled here to
3245 * However, as we check for exact result, doing it is safe.
3248 tarval_enable_fp_ops(1);
3249 tv = tarval_quo(get_mode_one(mode), tv);
3250 tarval_enable_fp_ops(rem);
3252 /* Do the transformation if the result is either exact or we are not
3253 using strict rules. */
3254 if (tv != tarval_bad &&
3255 (tarval_ieee754_get_exact() || (get_irg_fp_model(get_irn_irg(n)) & fp_strict_algebraic) == 0)) {
3256 ir_node *blk = get_nodes_block(n);
3257 ir_node *c = new_Const(tv);
3258 ir_node *a = get_Quot_left(n);
3259 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), blk, a, c, mode);
3260 ir_node *mem = get_Quot_mem(n);
3262 /* skip a potential Pin */
3263 mem = skip_Pin(mem);
3264 turn_into_tuple(n, pn_Quot_max);
3265 set_Tuple_pred(n, pn_Quot_M, mem);
3266 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(blk));
3267 set_Tuple_pred(n, pn_Quot_X_except, new_Bad());
3268 set_Tuple_pred(n, pn_Quot_res, m);
3269 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
3274 } /* transform_node_Quot */
3277 * Optimize Abs(x) into x if x is Confirmed >= 0
3278 * Optimize Abs(x) into -x if x is Confirmed <= 0
3279 * Optimize Abs(-x) int Abs(x)
3281 static ir_node *transform_node_Abs(ir_node *n)
3283 ir_node *c, *oldn = n;
3284 ir_node *a = get_Abs_op(n);
3287 HANDLE_UNOP_PHI(tarval_abs, a, c);
3289 switch (classify_value_sign(a)) {
3290 case value_classified_negative:
3291 mode = get_irn_mode(n);
3294 * We can replace the Abs by -x here.
3295 * We even could add a new Confirm here
3296 * (if not twos complement)
3298 * Note that -x would create a new node, so we could
3299 * not run it in the equivalent_node() context.
3301 n = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
3303 DBG_OPT_CONFIRM(oldn, n);
3305 case value_classified_positive:
3306 /* n is positive, Abs is not needed */
3309 DBG_OPT_CONFIRM(oldn, n);
3315 /* Abs(-x) = Abs(x) */
3316 mode = get_irn_mode(n);
3317 n = new_rd_Abs(get_irn_dbg_info(n), get_nodes_block(n), get_Minus_op(a), mode);
3318 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
3322 } /* transform_node_Abs */
3325 * Optimize -a CMP -b into b CMP a.
3326 * This works only for for modes where unary Minus
3328 * Note that two-complement integers can Overflow
3329 * so it will NOT work.
3331 * For == and != can be handled in Proj(Cmp)
3333 static ir_node *transform_node_Cmp(ir_node *n)
3336 ir_node *left = get_Cmp_left(n);
3337 ir_node *right = get_Cmp_right(n);
3339 if (is_Minus(left) && is_Minus(right) &&
3340 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
3341 ir_node *const new_left = get_Minus_op(right);
3342 ir_node *const new_right = get_Minus_op(left);
3343 n = new_rd_Cmp(get_irn_dbg_info(n), get_nodes_block(n), new_left, new_right);
3344 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP);
3347 } /* transform_node_Cmp */
3351 * Transform a Cond node.
3353 * Replace the Cond by a Jmp if it branches on a constant
3356 static ir_node *transform_node_Cond(ir_node *n)
3360 ir_node *a = get_Cond_selector(n);
3361 tarval *ta = value_of(a);
3362 ir_graph *irg = get_irn_irg(n);
3364 /* we need block info which is not available in floating irgs */
3365 if (get_irg_pinned(irg) == op_pin_state_floats)
3368 if ((ta != tarval_bad) &&
3369 (get_irn_mode(a) == mode_b) &&
3370 (get_opt_unreachable_code())) {
3371 /* It's a boolean Cond, branching on a boolean constant.
3372 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3373 ir_node *blk = get_nodes_block(n);
3374 jmp = new_r_Jmp(blk);
3375 turn_into_tuple(n, pn_Cond_max);
3376 if (ta == tarval_b_true) {
3377 set_Tuple_pred(n, pn_Cond_false, new_Bad());
3378 set_Tuple_pred(n, pn_Cond_true, jmp);
3380 set_Tuple_pred(n, pn_Cond_false, jmp);
3381 set_Tuple_pred(n, pn_Cond_true, new_Bad());
3383 /* We might generate an endless loop, so keep it alive. */
3384 add_End_keepalive(get_irg_end(irg), blk);
3387 } /* transform_node_Cond */
3390 * Prototype of a recursive transform function
3391 * for bitwise distributive transformations.
3393 typedef ir_node* (*recursive_transform)(ir_node *n);
3396 * makes use of distributive laws for and, or, eor
3397 * and(a OP c, b OP c) -> and(a, b) OP c
3398 * note, might return a different op than n
3400 static ir_node *transform_bitwise_distributive(ir_node *n,
3401 recursive_transform trans_func)
3404 ir_node *a = get_binop_left(n);
3405 ir_node *b = get_binop_right(n);
3406 ir_op *op = get_irn_op(a);
3407 ir_op *op_root = get_irn_op(n);
3409 if (op != get_irn_op(b))
3412 /* and(conv(a), conv(b)) -> conv(and(a,b)) */
3413 if (op == op_Conv) {
3414 ir_node *a_op = get_Conv_op(a);
3415 ir_node *b_op = get_Conv_op(b);
3416 ir_mode *a_mode = get_irn_mode(a_op);
3417 ir_mode *b_mode = get_irn_mode(b_op);
3418 if (a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3419 ir_node *blk = get_nodes_block(n);
3422 set_binop_left(n, a_op);
3423 set_binop_right(n, b_op);
3424 set_irn_mode(n, a_mode);
3426 n = new_r_Conv(blk, n, get_irn_mode(oldn));
3428 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
3434 /* nothing to gain here */
3438 if (op == op_Shrs || op == op_Shr || op == op_Shl
3439 || op == op_And || op == op_Or || op == op_Eor) {
3440 ir_node *a_left = get_binop_left(a);
3441 ir_node *a_right = get_binop_right(a);
3442 ir_node *b_left = get_binop_left(b);
3443 ir_node *b_right = get_binop_right(b);
3445 ir_node *op1 = NULL;
3446 ir_node *op2 = NULL;
3448 if (is_op_commutative(op)) {
3449 if (a_left == b_left) {
3453 } else if (a_left == b_right) {
3457 } else if (a_right == b_left) {
3463 if (a_right == b_right) {
3470 /* (a sop c) & (b sop c) => (a & b) sop c */
3471 ir_node *blk = get_nodes_block(n);
3473 ir_node *new_n = exact_copy(n);
3474 set_binop_left(new_n, op1);
3475 set_binop_right(new_n, op2);
3476 new_n = trans_func(new_n);
3478 if (op_root == op_Eor && op == op_Or) {
3479 dbg_info *dbgi = get_irn_dbg_info(n);
3480 ir_mode *mode = get_irn_mode(c);
3482 c = new_rd_Not(dbgi, blk, c, mode);
3483 n = new_rd_And(dbgi, blk, new_n, c, mode);
3486 set_nodes_block(n, blk);
3487 set_binop_left(n, new_n);
3488 set_binop_right(n, c);
3492 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3503 static ir_node *transform_node_And(ir_node *n)
3505 ir_node *c, *oldn = n;
3506 ir_node *a = get_And_left(n);
3507 ir_node *b = get_And_right(n);
3509 vrp_attr *a_vrp, *b_vrp;
3511 mode = get_irn_mode(n);
3512 HANDLE_BINOP_PHI((eval_func) tarval_and, a, b, c, mode);
3514 /* we can evaluate 2 Projs of the same Cmp */
3515 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3516 ir_node *pred_a = get_Proj_pred(a);
3517 ir_node *pred_b = get_Proj_pred(b);
3518 if (pred_a == pred_b) {
3519 dbg_info *dbgi = get_irn_dbg_info(n);
3520 pn_Cmp pn_a = get_Proj_proj(a);
3521 pn_Cmp pn_b = get_Proj_proj(b);
3522 /* yes, we can simply calculate with pncs */
3523 pn_Cmp new_pnc = pn_a & pn_b;
3525 return new_rd_Proj(dbgi, pred_a, mode_b, new_pnc);
3530 ir_node *op = get_Not_op(b);
3532 ir_node *ba = get_And_left(op);
3533 ir_node *bb = get_And_right(op);
3535 /* it's enough to test the following cases due to normalization! */
3536 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3537 /* (a|b) & ~(a&b) = a^b */
3538 ir_node *block = get_nodes_block(n);
3540 n = new_rd_Eor(get_irn_dbg_info(n), block, ba, bb, mode);
3541 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3549 ir_node *op = get_Not_op(a);
3551 ir_node *aa = get_And_left(op);
3552 ir_node *ab = get_And_right(op);
3554 /* it's enough to test the following cases due to normalization! */
3555 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3556 /* (a|b) & ~(a&b) = a^b */
3557 ir_node *block = get_nodes_block(n);
3559 n = new_rd_Eor(get_irn_dbg_info(n), block, aa, ab, mode);
3560 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3567 ir_node *al = get_Eor_left(a);
3568 ir_node *ar = get_Eor_right(a);
3571 /* (b ^ a) & b -> ~a & b */
3572 dbg_info *dbg = get_irn_dbg_info(n);
3573 ir_node *block = get_nodes_block(n);
3575 ar = new_rd_Not(dbg, block, ar, mode);
3576 n = new_rd_And(dbg, block, ar, b, mode);
3577 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3581 /* (a ^ b) & b -> ~a & b */
3582 dbg_info *dbg = get_irn_dbg_info(n);
3583 ir_node *block = get_nodes_block(n);
3585 al = new_rd_Not(dbg, block, al, mode);
3586 n = new_rd_And(dbg, block, al, b, mode);
3587 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3592 ir_node *bl = get_Eor_left(b);
3593 ir_node *br = get_Eor_right(b);
3596 /* a & (a ^ b) -> a & ~b */
3597 dbg_info *dbg = get_irn_dbg_info(n);
3598 ir_node *block = get_nodes_block(n);
3600 br = new_rd_Not(dbg, block, br, mode);
3601 n = new_rd_And(dbg, block, br, a, mode);
3602 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3606 /* a & (b ^ a) -> a & ~b */
3607 dbg_info *dbg = get_irn_dbg_info(n);
3608 ir_node *block = get_nodes_block(n);
3610 bl = new_rd_Not(dbg, block, bl, mode);
3611 n = new_rd_And(dbg, block, bl, a, mode);
3612 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3616 if (is_Not(a) && is_Not(b)) {
3617 /* ~a & ~b = ~(a|b) */
3618 ir_node *block = get_nodes_block(n);
3619 ir_mode *mode = get_irn_mode(n);
3623 n = new_rd_Or(get_irn_dbg_info(n), block, a, b, mode);
3624 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
3625 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3629 b_vrp = vrp_get_info(b);
3630 if (is_Const(a) && b_vrp && (tarval_cmp(tarval_or(get_Const_tarval(a),
3631 b_vrp->bits_not_set), get_Const_tarval(a)) == pn_Cmp_Eq)) {
3637 a_vrp = vrp_get_info(a);
3638 if (is_Const(b) && a_vrp && (tarval_cmp(tarval_or(get_Const_tarval(b),
3639 a_vrp->bits_not_set), get_Const_tarval(b)) == pn_Cmp_Eq)) {
3643 n = transform_bitwise_distributive(n, transform_node_And);
3646 } /* transform_node_And */
3651 static ir_node *transform_node_Eor(ir_node *n)
3653 ir_node *c, *oldn = n;
3654 ir_node *a = get_Eor_left(n);
3655 ir_node *b = get_Eor_right(n);
3656 ir_mode *mode = get_irn_mode(n);
3658 HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode);
3660 /* we can evaluate 2 Projs of the same Cmp */
3661 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3662 ir_node *pred_a = get_Proj_pred(a);
3663 ir_node *pred_b = get_Proj_pred(b);
3664 if (pred_a == pred_b) {
3665 dbg_info *dbgi = get_irn_dbg_info(n);
3666 pn_Cmp pn_a = get_Proj_proj(a);
3667 pn_Cmp pn_b = get_Proj_proj(b);
3668 /* yes, we can simply calculate with pncs */
3669 pn_Cmp new_pnc = pn_a ^ pn_b;
3671 return new_rd_Proj(dbgi, pred_a, mode_b, new_pnc);
3676 ir_graph *irg = get_irn_irg(n);
3678 n = new_rd_Const(get_irn_dbg_info(n), irg, get_mode_null(mode));
3679 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3683 /* normalize not nodes... ~a ^ b => a ^ ~b */
3685 dbg_info *dbg = get_irn_dbg_info(n);
3686 ir_node *block = get_nodes_block(n);
3687 ir_node *new_not = new_rd_Not(dbg, block, b, mode);
3688 ir_node *new_left = get_Not_op(a);
3689 n = new_rd_Eor(dbg, block, new_left, new_not, mode);
3690 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3694 /* x ^ 1...1 -> ~1 */
3695 if (is_Const(b) && is_Const_all_one(b)) {
3696 n = new_r_Not(get_nodes_block(n), a, mode);
3697 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3701 n = transform_bitwise_distributive(n, transform_node_Eor);
3703 } /* transform_node_Eor */
3708 static ir_node *transform_node_Not(ir_node *n)
3710 ir_node *c, *oldn = n;
3711 ir_node *a = get_Not_op(n);
3712 ir_mode *mode = get_irn_mode(n);
3714 HANDLE_UNOP_PHI(tarval_not,a,c);
3716 /* check for a boolean Not */
3717 if (mode == mode_b && is_Proj(a)) {
3718 ir_node *a_pred = get_Proj_pred(a);
3719 if (is_Cmp(a_pred)) {
3720 /* We negate a Cmp. The Cmp has the negated result anyways! */
3721 n = new_r_Proj(get_Proj_pred(a),
3722 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3723 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3728 /* normalize ~(a ^ b) => a ^ ~b */
3730 dbg_info *dbg = get_irn_dbg_info(n);
3731 ir_node *block = get_nodes_block(n);
3732 ir_node *eor_right = get_Eor_right(a);
3733 ir_node *eor_left = get_Eor_left(a);
3734 eor_right = new_rd_Not(dbg, block, eor_right, mode);
3735 n = new_rd_Eor(dbg, block, eor_left, eor_right, mode);
3739 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3740 if (is_Minus(a)) { /* ~-x -> x + -1 */
3741 dbg_info *dbg = get_irn_dbg_info(n);
3742 ir_graph *irg = get_irn_irg(n);
3743 ir_node *block = get_nodes_block(n);
3744 ir_node *add_l = get_Minus_op(a);
3745 ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode));
3746 n = new_rd_Add(dbg, block, add_l, add_r, mode);
3747 } else if (is_Add(a)) {
3748 ir_node *add_r = get_Add_right(a);
3749 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3750 /* ~(x + -1) = -x */
3751 ir_node *op = get_Add_left(a);
3752 ir_node *blk = get_nodes_block(n);
3753 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, get_irn_mode(n));
3754 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3759 } /* transform_node_Not */
3762 * Transform a Minus.
3766 * -(a >>u (size-1)) = a >>s (size-1)
3767 * -(a >>s (size-1)) = a >>u (size-1)
3768 * -(a * const) -> a * -const
3770 static ir_node *transform_node_Minus(ir_node *n)
3772 ir_node *c, *oldn = n;
3773 ir_node *a = get_Minus_op(n);
3776 HANDLE_UNOP_PHI(tarval_neg,a,c);
3778 mode = get_irn_mode(a);
3779 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3780 /* the following rules are only to twos-complement */
3783 ir_node *op = get_Not_op(a);
3784 tarval *tv = get_mode_one(mode);
3785 ir_node *blk = get_nodes_block(n);
3786 ir_node *c = new_Const(tv);
3787 n = new_rd_Add(get_irn_dbg_info(n), blk, op, c, mode);
3788 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3792 ir_node *c = get_Shr_right(a);
3795 tarval *tv = get_Const_tarval(c);
3797 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3798 /* -(a >>u (size-1)) = a >>s (size-1) */
3799 ir_node *v = get_Shr_left(a);
3801 n = new_rd_Shrs(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3802 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3808 ir_node *c = get_Shrs_right(a);
3811 tarval *tv = get_Const_tarval(c);
3813 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3814 /* -(a >>s (size-1)) = a >>u (size-1) */
3815 ir_node *v = get_Shrs_left(a);
3817 n = new_rd_Shr(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3818 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3825 /* - (a-b) = b - a */
3826 ir_node *la = get_Sub_left(a);
3827 ir_node *ra = get_Sub_right(a);
3828 ir_node *blk = get_nodes_block(n);
3830 n = new_rd_Sub(get_irn_dbg_info(n), blk, ra, la, mode);
3831 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3835 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3836 ir_node *mul_l = get_Mul_left(a);
3837 ir_node *mul_r = get_Mul_right(a);
3838 tarval *tv = value_of(mul_r);
3839 if (tv != tarval_bad) {
3840 tv = tarval_neg(tv);
3841 if (tv != tarval_bad) {
3842 ir_node *cnst = new_Const(tv);
3843 dbg_info *dbg = get_irn_dbg_info(a);
3844 ir_node *block = get_nodes_block(a);
3845 n = new_rd_Mul(dbg, block, mul_l, cnst, mode);
3846 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3853 } /* transform_node_Minus */
3856 * Transform a Cast_type(Const) into a new Const_type
3858 static ir_node *transform_node_Cast(ir_node *n)
3861 ir_node *pred = get_Cast_op(n);
3862 ir_type *tp = get_irn_type(n);
3864 if (is_Const(pred) && get_Const_type(pred) != tp) {
3865 ir_graph *irg = get_irn_irg(n);
3866 n = new_rd_Const_type(NULL, irg, get_Const_tarval(pred), tp);
3867 DBG_OPT_CSTEVAL(oldn, n);
3868 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3869 ir_graph *irg = get_irn_irg(n);
3870 n = new_rd_SymConst_type(NULL, irg, get_irn_mode(pred),
3871 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3872 DBG_OPT_CSTEVAL(oldn, n);
3876 } /* transform_node_Cast */
3879 * Transform a Proj(Load) with a non-null address.
3881 static ir_node *transform_node_Proj_Load(ir_node *proj)
3883 if (get_opt_ldst_only_null_ptr_exceptions()) {
3884 if (get_irn_mode(proj) == mode_X) {
3885 ir_node *load = get_Proj_pred(proj);
3887 /* get the Load address */
3888 const ir_node *addr = get_Load_ptr(load);
3889 const ir_node *confirm;
3891 if (value_not_null(addr, &confirm)) {
3892 if (confirm == NULL) {
3893 /* this node may float if it did not depend on a Confirm */
3894 set_irn_pinned(load, op_pin_state_floats);
3896 if (get_Proj_proj(proj) == pn_Load_X_except) {
3897 ir_graph *irg = get_irn_irg(proj);
3898 DBG_OPT_EXC_REM(proj);
3899 return get_irg_bad(irg);
3901 ir_node *blk = get_nodes_block(load);
3902 return new_r_Jmp(blk);
3908 } /* transform_node_Proj_Load */
3911 * Transform a Proj(Store) with a non-null address.
3913 static ir_node *transform_node_Proj_Store(ir_node *proj)
3915 if (get_opt_ldst_only_null_ptr_exceptions()) {
3916 if (get_irn_mode(proj) == mode_X) {
3917 ir_node *store = get_Proj_pred(proj);
3919 /* get the load/store address */
3920 const ir_node *addr = get_Store_ptr(store);
3921 const ir_node *confirm;
3923 if (value_not_null(addr, &confirm)) {
3924 if (confirm == NULL) {
3925 /* this node may float if it did not depend on a Confirm */
3926 set_irn_pinned(store, op_pin_state_floats);
3928 if (get_Proj_proj(proj) == pn_Store_X_except) {
3929 ir_graph *irg = get_irn_irg(proj);
3930 DBG_OPT_EXC_REM(proj);
3931 return get_irg_bad(irg);
3933 ir_node *blk = get_nodes_block(store);
3934 return new_r_Jmp(blk);
3940 } /* transform_node_Proj_Store */
3943 * Transform a Proj(Div) with a non-zero value.
3944 * Removes the exceptions and routes the memory to the NoMem node.
3946 static ir_node *transform_node_Proj_Div(ir_node *proj)
3948 ir_node *div = get_Proj_pred(proj);
3949 ir_node *b = get_Div_right(div);
3950 ir_node *res, *new_mem;
3951 const ir_node *confirm;
3954 if (value_not_zero(b, &confirm)) {
3955 /* div(x, y) && y != 0 */
3956 if (confirm == NULL) {
3957 /* we are sure we have a Const != 0 */
3958 new_mem = get_Div_mem(div);
3959 new_mem = skip_Pin(new_mem);
3960 set_Div_mem(div, new_mem);
3961 set_irn_pinned(div, op_pin_state_floats);
3964 proj_nr = get_Proj_proj(proj);
3966 case pn_Div_X_regular:
3967 return new_r_Jmp(get_nodes_block(div));
3969 case pn_Div_X_except:
3970 /* we found an exception handler, remove it */
3971 DBG_OPT_EXC_REM(proj);
3975 ir_graph *irg = get_irn_irg(proj);
3976 res = get_Div_mem(div);
3977 new_mem = get_irg_no_mem(irg);
3980 /* This node can only float up to the Confirm block */
3981 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3983 set_irn_pinned(div, op_pin_state_floats);
3984 /* this is a Div without exception, we can remove the memory edge */
3985 set_Div_mem(div, new_mem);
3991 } /* transform_node_Proj_Div */
3994 * Transform a Proj(Mod) with a non-zero value.
3995 * Removes the exceptions and routes the memory to the NoMem node.
3997 static ir_node *transform_node_Proj_Mod(ir_node *proj)
3999 ir_node *mod = get_Proj_pred(proj);
4000 ir_node *b = get_Mod_right(mod);
4001 ir_node *res, *new_mem;
4002 const ir_node *confirm;
4005 if (value_not_zero(b, &confirm)) {
4006 /* mod(x, y) && y != 0 */
4007 proj_nr = get_Proj_proj(proj);
4009 if (confirm == NULL) {
4010 /* we are sure we have a Const != 0 */
4011 new_mem = get_Mod_mem(mod);
4012 new_mem = skip_Pin(new_mem);
4013 set_Mod_mem(mod, new_mem);
4014 set_irn_pinned(mod, op_pin_state_floats);
4019 case pn_Mod_X_regular:
4020 return new_r_Jmp(get_irn_n(mod, -1));
4022 case pn_Mod_X_except:
4023 /* we found an exception handler, remove it */
4024 DBG_OPT_EXC_REM(proj);
4028 ir_graph *irg = get_irn_irg(proj);
4029 res = get_Mod_mem(mod);
4030 new_mem = get_irg_no_mem(irg);
4033 /* This node can only float up to the Confirm block */
4034 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
4036 /* this is a Mod without exception, we can remove the memory edge */
4037 set_Mod_mem(mod, new_mem);
4041 if (get_Mod_left(mod) == b) {
4042 /* a % a = 0 if a != 0 */
4043 ir_mode *mode = get_irn_mode(proj);
4044 ir_node *res = new_Const(get_mode_null(mode));
4046 DBG_OPT_CSTEVAL(mod, res);
4052 } /* transform_node_Proj_Mod */
4055 * Transform a Proj(DivMod) with a non-zero value.
4056 * Removes the exceptions and routes the memory to the NoMem node.
4058 static ir_node *transform_node_Proj_DivMod(ir_node *proj)
4060 ir_node *divmod = get_Proj_pred(proj);
4061 ir_node *b = get_DivMod_right(divmod);
4062 ir_node *res, *new_mem;
4063 const ir_node *confirm;
4066 if (value_not_zero(b, &confirm)) {
4067 /* DivMod(x, y) && y != 0 */
4068 proj_nr = get_Proj_proj(proj);
4070 if (confirm == NULL) {
4071 /* we are sure we have a Const != 0 */
4072 new_mem = get_DivMod_mem(divmod);
4073 new_mem = skip_Pin(new_mem);
4074 set_DivMod_mem(divmod, new_mem);
4075 set_irn_pinned(divmod, op_pin_state_floats);
4080 case pn_DivMod_X_regular:
4081 return new_r_Jmp(get_nodes_block(divmod));
4083 case pn_DivMod_X_except:
4084 /* we found an exception handler, remove it */
4085 DBG_OPT_EXC_REM(proj);
4089 ir_graph *irg = get_irn_irg(proj);
4090 res = get_DivMod_mem(divmod);
4091 new_mem = get_irg_no_mem(irg);
4094 /* This node can only float up to the Confirm block */
4095 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
4097 /* this is a DivMod without exception, we can remove the memory edge */
4098 set_DivMod_mem(divmod, new_mem);
4102 case pn_DivMod_res_mod:
4103 if (get_DivMod_left(divmod) == b) {
4104 /* a % a = 0 if a != 0 */
4105 ir_mode *mode = get_irn_mode(proj);
4106 ir_node *res = new_Const(get_mode_null(mode));
4108 DBG_OPT_CSTEVAL(divmod, res);
4114 } /* transform_node_Proj_DivMod */
4117 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
4119 static ir_node *transform_node_Proj_Cond(ir_node *proj)
4121 if (get_opt_unreachable_code()) {
4122 ir_node *n = get_Proj_pred(proj);
4123 ir_node *b = get_Cond_selector(n);
4125 if (mode_is_int(get_irn_mode(b))) {
4126 tarval *tb = value_of(b);
4128 if (tb != tarval_bad) {
4129 /* we have a constant switch */
4130 long num = get_Proj_proj(proj);
4132 if (num != get_Cond_default_proj(n)) { /* we cannot optimize default Proj's yet */
4133 if (get_tarval_long(tb) == num) {
4134 /* Do NOT create a jump here, or we will have 2 control flow ops
4135 * in a block. This case is optimized away in optimize_cf(). */
4138 ir_graph *irg = get_irn_irg(proj);
4139 /* this case will NEVER be taken, kill it */
4140 return get_irg_bad(irg);
4144 long num = get_Proj_proj(proj);
4145 vrp_attr *b_vrp = vrp_get_info(b);
4146 if (num != get_Cond_default_proj(n) && b_vrp) {
4147 /* Try handling with vrp data. We only remove dead parts. */
4148 tarval *tp = new_tarval_from_long(num, get_irn_mode(b));
4150 if (b_vrp->range_type == VRP_RANGE) {
4151 pn_Cmp cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
4152 pn_Cmp cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
4154 if ((cmp_result & pn_Cmp_Gt) == cmp_result && (cmp_result2
4155 & pn_Cmp_Lt) == cmp_result2) {
4156 ir_graph *irg = get_irn_irg(proj);
4157 return get_irg_bad(irg);
4159 } else if (b_vrp->range_type == VRP_ANTIRANGE) {
4160 pn_Cmp cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
4161 pn_Cmp cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
4163 if ((cmp_result & pn_Cmp_Le) == cmp_result && (cmp_result2
4164 & pn_Cmp_Ge) == cmp_result2) {
4165 ir_graph *irg = get_irn_irg(proj);
4166 return get_irg_bad(irg);
4171 tarval_and( b_vrp->bits_set, tp),
4174 ir_graph *irg = get_irn_irg(proj);
4175 return get_irg_bad(irg);
4181 tarval_not(b_vrp->bits_not_set)),
4182 tarval_not(b_vrp->bits_not_set))
4184 ir_graph *irg = get_irn_irg(proj);
4185 return get_irg_bad(irg);
4194 } /* transform_node_Proj_Cond */
4197 * Create a 0 constant of given mode.
4199 static ir_node *create_zero_const(ir_mode *mode)
4201 tarval *tv = get_mode_null(mode);
4202 ir_node *cnst = new_Const(tv);
4207 /* the order of the values is important! */
4208 typedef enum const_class {
4214 static const_class classify_const(const ir_node* n)
4216 if (is_Const(n)) return const_const;
4217 if (is_irn_constlike(n)) return const_like;
4222 * Determines whether r is more constlike or has a larger index (in that order)
4225 static int operands_are_normalized(const ir_node *l, const ir_node *r)
4227 const const_class l_order = classify_const(l);
4228 const const_class r_order = classify_const(r);
4230 l_order > r_order ||
4231 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
4235 * Normalizes and optimizes Cmp nodes.
4237 static ir_node *transform_node_Proj_Cmp(ir_node *proj)
4239 ir_node *n = get_Proj_pred(proj);
4240 ir_node *left = get_Cmp_left(n);
4241 ir_node *right = get_Cmp_right(n);
4244 ir_mode *mode = NULL;
4245 long proj_nr = get_Proj_proj(proj);
4247 /* we can evaluate some cases directly */
4250 return new_Const(get_tarval_b_false());
4252 return new_Const(get_tarval_b_true());
4254 if (!mode_is_float(get_irn_mode(left)))
4255 return new_Const(get_tarval_b_true());
4261 /* remove Casts of both sides */
4262 left = skip_Cast(left);
4263 right = skip_Cast(right);
4265 /* Remove unnecessary conversions */
4266 /* TODO handle constants */
4267 if (is_Conv(left) && is_Conv(right)) {
4268 ir_mode *mode = get_irn_mode(left);
4269 ir_node *op_left = get_Conv_op(left);
4270 ir_node *op_right = get_Conv_op(right);
4271 ir_mode *mode_left = get_irn_mode(op_left);
4272 ir_mode *mode_right = get_irn_mode(op_right);
4274 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
4275 && mode_left != mode_b && mode_right != mode_b) {
4276 ir_node *block = get_nodes_block(n);
4278 if (mode_left == mode_right) {
4282 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
4283 } else if (smaller_mode(mode_left, mode_right)) {
4284 left = new_r_Conv(block, op_left, mode_right);
4287 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4288 } else if (smaller_mode(mode_right, mode_left)) {
4290 right = new_r_Conv(block, op_right, mode_left);
4292 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4297 /* remove operation on both sides if possible */
4298 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4300 * The following operations are NOT safe for floating point operations, for instance
4301 * 1.0 + inf == 2.0 + inf, =/=> x == y
4303 if (mode_is_int(get_irn_mode(left))) {
4304 unsigned lop = get_irn_opcode(left);
4306 if (lop == get_irn_opcode(right)) {
4307 ir_node *ll, *lr, *rl, *rr;
4309 /* same operation on both sides, try to remove */
4313 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
4314 left = get_unop_op(left);
4315 right = get_unop_op(right);
4317 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4320 ll = get_Add_left(left);
4321 lr = get_Add_right(left);
4322 rl = get_Add_left(right);
4323 rr = get_Add_right(right);
4326 /* X + a CMP X + b ==> a CMP b */
4330 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4331 } else if (ll == rr) {
4332 /* X + a CMP b + X ==> a CMP b */
4336 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4337 } else if (lr == rl) {
4338 /* a + X CMP X + b ==> a CMP b */
4342 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4343 } else if (lr == rr) {
4344 /* a + X CMP b + X ==> a CMP b */
4348 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4352 ll = get_Sub_left(left);
4353 lr = get_Sub_right(left);
4354 rl = get_Sub_left(right);
4355 rr = get_Sub_right(right);
4358 /* X - a CMP X - b ==> a CMP b */
4362 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4363 } else if (lr == rr) {
4364 /* a - X CMP b - X ==> a CMP b */
4368 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4372 if (get_Rotl_right(left) == get_Rotl_right(right)) {
4373 /* a ROTL X CMP b ROTL X ==> a CMP b */
4374 left = get_Rotl_left(left);
4375 right = get_Rotl_left(right);
4377 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4385 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4386 if (is_Add(left) || is_Sub(left)) {
4387 ir_node *ll = get_binop_left(left);
4388 ir_node *lr = get_binop_right(left);
4390 if (lr == right && is_Add(left)) {
4397 right = create_zero_const(get_irn_mode(left));
4399 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4402 if (is_Add(right) || is_Sub(right)) {
4403 ir_node *rl = get_binop_left(right);
4404 ir_node *rr = get_binop_right(right);
4406 if (rr == left && is_Add(right)) {
4413 right = create_zero_const(get_irn_mode(left));
4415 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4418 if (is_And(left) && is_Const(right)) {
4419 ir_node *ll = get_binop_left(left);
4420 ir_node *lr = get_binop_right(left);
4421 if (is_Shr(ll) && is_Const(lr)) {
4422 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
4423 ir_node *block = get_nodes_block(n);
4424 ir_mode *mode = get_irn_mode(left);
4426 ir_node *llr = get_Shr_right(ll);
4427 if (is_Const(llr)) {
4428 dbg_info *dbg = get_irn_dbg_info(left);
4430 tarval *c1 = get_Const_tarval(llr);
4431 tarval *c2 = get_Const_tarval(lr);
4432 tarval *c3 = get_Const_tarval(right);
4433 tarval *mask = tarval_shl(c2, c1);
4434 tarval *value = tarval_shl(c3, c1);
4436 left = new_rd_And(dbg, block, get_Shr_left(ll), new_Const(mask), mode);
4437 right = new_Const(value);
4442 } /* mode_is_int(...) */
4443 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
4445 /* replace mode_b compares with ands/ors */
4446 if (get_irn_mode(left) == mode_b) {
4447 ir_node *block = get_nodes_block(n);
4451 case pn_Cmp_Le: bres = new_r_Or( block, new_r_Not(block, left, mode_b), right, mode_b); break;
4452 case pn_Cmp_Lt: bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b); break;
4453 case pn_Cmp_Ge: bres = new_r_Or( block, left, new_r_Not(block, right, mode_b), mode_b); break;
4454 case pn_Cmp_Gt: bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b); break;
4455 case pn_Cmp_Lg: bres = new_r_Eor(block, left, right, mode_b); break;
4456 case pn_Cmp_Eq: bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b); break;
4457 default: bres = NULL;
4460 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4466 * First step: normalize the compare op
4467 * by placing the constant on the right side
4468 * or moving the lower address node to the left.
4470 if (!operands_are_normalized(left, right)) {
4476 proj_nr = get_inversed_pnc(proj_nr);
4481 * Second step: Try to reduce the magnitude
4482 * of a constant. This may help to generate better code
4483 * later and may help to normalize more compares.
4484 * Of course this is only possible for integer values.
4486 tv = value_of(right);
4487 if (tv != tarval_bad) {
4488 mode = get_irn_mode(right);
4490 /* TODO extend to arbitrary constants */
4491 if (is_Conv(left) && tarval_is_null(tv)) {
4492 ir_node *op = get_Conv_op(left);
4493 ir_mode *op_mode = get_irn_mode(op);
4496 * UpConv(x) REL 0 ==> x REL 0
4497 * Don't do this for float values as it's unclear whether it is a
4498 * win. (on the other side it makes detection/creation of fabs hard)
4500 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4501 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
4502 mode_is_signed(mode) || !mode_is_signed(op_mode)) &&
4503 !mode_is_float(mode)) {
4504 tv = get_mode_null(op_mode);
4508 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4512 if (tv != tarval_bad) {
4513 /* the following optimization is possible on modes without Overflow
4514 * on Unary Minus or on == and !=:
4515 * -a CMP c ==> a swap(CMP) -c
4517 * Beware: for two-complement Overflow may occur, so only == and != can
4518 * be optimized, see this:
4519 * -MININT < 0 =/=> MININT > 0 !!!
4521 if (is_Minus(left) &&
4522 (!mode_overflow_on_unary_Minus(mode) ||
4523 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
4524 tv = tarval_neg(tv);
4526 if (tv != tarval_bad) {
4527 left = get_Minus_op(left);
4528 proj_nr = get_inversed_pnc(proj_nr);
4530 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4532 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
4533 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4534 tv = tarval_not(tv);
4536 if (tv != tarval_bad) {
4537 left = get_Not_op(left);
4539 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4543 /* for integer modes, we have more */
4544 if (mode_is_int(mode)) {
4545 /* Ne includes Unordered which is not possible on integers.
4546 * However, frontends often use this wrong, so fix it here */
4547 if (proj_nr & pn_Cmp_Uo) {
4548 proj_nr &= ~pn_Cmp_Uo;
4549 set_Proj_proj(proj, proj_nr);
4552 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4553 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
4554 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
4555 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4557 if (tv != tarval_bad) {
4558 proj_nr ^= pn_Cmp_Eq;
4560 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4563 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4564 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
4565 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
4566 tv = tarval_add(tv, get_mode_one(mode));
4568 if (tv != tarval_bad) {
4569 proj_nr ^= pn_Cmp_Eq;
4571 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4575 /* the following reassociations work only for == and != */
4576 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4578 #if 0 /* Might be not that good in general */
4579 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4580 if (tarval_is_null(tv) && is_Sub(left)) {
4581 right = get_Sub_right(left);
4582 left = get_Sub_left(left);
4584 tv = value_of(right);
4586 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4590 if (tv != tarval_bad) {
4591 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4593 ir_node *c1 = get_Sub_right(left);
4594 tarval *tv2 = value_of(c1);
4596 if (tv2 != tarval_bad) {
4597 tv2 = tarval_add(tv, value_of(c1));
4599 if (tv2 != tarval_bad) {
4600 left = get_Sub_left(left);
4603 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4607 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4608 else if (is_Add(left)) {
4609 ir_node *a_l = get_Add_left(left);
4610 ir_node *a_r = get_Add_right(left);
4614 if (is_Const(a_l)) {
4616 tv2 = value_of(a_l);
4619 tv2 = value_of(a_r);
4622 if (tv2 != tarval_bad) {
4623 tv2 = tarval_sub(tv, tv2, NULL);
4625 if (tv2 != tarval_bad) {
4629 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4633 /* -a == c ==> a == -c, -a != c ==> a != -c */
4634 else if (is_Minus(left)) {
4635 tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4637 if (tv2 != tarval_bad) {
4638 left = get_Minus_op(left);
4641 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4646 /* the following reassociations work only for <= */
4647 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4648 if (tv != tarval_bad) {
4649 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4650 if (is_Abs(left)) { // TODO something is missing here
4656 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4657 switch (get_irn_opcode(left)) {
4661 c1 = get_And_right(left);
4664 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4665 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4667 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4669 /* TODO: move to constant evaluation */
4670 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4672 DBG_OPT_CSTEVAL(proj, c1);
4676 if (tarval_is_single_bit(tv)) {
4678 * optimization for AND:
4680 * And(x, C) == C ==> And(x, C) != 0
4681 * And(x, C) != C ==> And(X, C) == 0
4683 * if C is a single Bit constant.
4686 /* check for Constant's match. We have check hare the tarvals,
4687 because our const might be changed */
4688 if (get_Const_tarval(c1) == tv) {
4689 /* fine: do the transformation */
4690 tv = get_mode_null(get_tarval_mode(tv));
4691 proj_nr ^= pn_Cmp_Leg;
4693 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4699 c1 = get_Or_right(left);
4700 if (is_Const(c1) && tarval_is_null(tv)) {
4702 * Or(x, C) == 0 && C != 0 ==> FALSE
4703 * Or(x, C) != 0 && C != 0 ==> TRUE
4705 if (! tarval_is_null(get_Const_tarval(c1))) {
4706 /* TODO: move to constant evaluation */
4707 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4709 DBG_OPT_CSTEVAL(proj, c1);
4716 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4718 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4721 c1 = get_Shl_right(left);
4723 tarval *tv1 = get_Const_tarval(c1);
4724 ir_mode *mode = get_irn_mode(left);
4725 tarval *minus1 = get_mode_all_one(mode);
4726 tarval *amask = tarval_shr(minus1, tv1);
4727 tarval *cmask = tarval_shl(minus1, tv1);
4730 if (tarval_and(tv, cmask) != tv) {
4731 /* condition not met */
4732 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4734 DBG_OPT_CSTEVAL(proj, c1);
4737 sl = get_Shl_left(left);
4738 blk = get_nodes_block(n);
4739 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_Const(amask), mode);
4740 tv = tarval_shr(tv, tv1);
4742 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4747 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4749 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4752 c1 = get_Shr_right(left);
4754 tarval *tv1 = get_Const_tarval(c1);
4755 ir_mode *mode = get_irn_mode(left);
4756 tarval *minus1 = get_mode_all_one(mode);
4757 tarval *amask = tarval_shl(minus1, tv1);
4758 tarval *cmask = tarval_shr(minus1, tv1);
4761 if (tarval_and(tv, cmask) != tv) {
4762 /* condition not met */
4763 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4765 DBG_OPT_CSTEVAL(proj, c1);
4768 sl = get_Shr_left(left);
4769 blk = get_nodes_block(n);
4770 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_Const(amask), mode);
4771 tv = tarval_shl(tv, tv1);
4773 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4778 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4780 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4783 c1 = get_Shrs_right(left);
4785 tarval *tv1 = get_Const_tarval(c1);
4786 ir_mode *mode = get_irn_mode(left);
4787 tarval *minus1 = get_mode_all_one(mode);
4788 tarval *amask = tarval_shl(minus1, tv1);
4789 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4792 cond = tarval_sub(cond, tv1, NULL);
4793 cond = tarval_shrs(tv, cond);
4795 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4796 /* condition not met */
4797 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4799 DBG_OPT_CSTEVAL(proj, c1);
4802 sl = get_Shrs_left(left);
4803 blk = get_nodes_block(n);
4804 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_Const(amask), mode);
4805 tv = tarval_shl(tv, tv1);
4807 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4812 } /* tarval != bad */
4815 if (changed & 2) /* need a new Const */
4816 right = new_Const(tv);
4818 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4819 ir_node *op = get_Proj_pred(left);
4821 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4822 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4823 ir_node *c = get_binop_right(op);
4826 tarval *tv = get_Const_tarval(c);
4828 if (tarval_is_single_bit(tv)) {
4829 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4830 ir_node *v = get_binop_left(op);
4831 ir_node *blk = get_irn_n(op, -1);
4832 ir_mode *mode = get_irn_mode(v);
4834 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4835 left = new_rd_And(get_irn_dbg_info(op), blk, v, new_Const(tv), mode);
4837 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4844 ir_node *block = get_nodes_block(n);
4846 /* create a new compare */
4847 n = new_rd_Cmp(get_irn_dbg_info(n), block, left, right);
4848 proj = new_rd_Proj(get_irn_dbg_info(proj), n, get_irn_mode(proj), proj_nr);
4852 } /* transform_node_Proj_Cmp */
4855 * Optimize CopyB(mem, x, x) into a Nop.
4857 static ir_node *transform_node_Proj_CopyB(ir_node *proj)
4859 ir_node *copyb = get_Proj_pred(proj);
4860 ir_node *a = get_CopyB_dst(copyb);
4861 ir_node *b = get_CopyB_src(copyb);
4864 switch (get_Proj_proj(proj)) {
4865 case pn_CopyB_X_regular:
4866 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4867 DBG_OPT_EXC_REM(proj);
4868 proj = new_r_Jmp(get_nodes_block(copyb));
4870 case pn_CopyB_X_except:
4871 DBG_OPT_EXC_REM(proj);
4872 proj = get_irg_bad(get_irn_irg(proj));
4879 } /* transform_node_Proj_CopyB */
4882 * Optimize Bounds(idx, idx, upper) into idx.
4884 static ir_node *transform_node_Proj_Bound(ir_node *proj)
4886 ir_node *oldn = proj;
4887 ir_node *bound = get_Proj_pred(proj);
4888 ir_node *idx = get_Bound_index(bound);
4889 ir_node *pred = skip_Proj(idx);
4892 if (idx == get_Bound_lower(bound))
4894 else if (is_Bound(pred)) {
4896 * idx was Bounds checked in the same MacroBlock previously,
4897 * it is still valid if lower <= pred_lower && pred_upper <= upper.
4899 ir_node *lower = get_Bound_lower(bound);
4900 ir_node *upper = get_Bound_upper(bound);
4901 if (get_Bound_lower(pred) == lower &&
4902 get_Bound_upper(pred) == upper &&
4903 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
4905 * One could expect that we simply return the previous
4906 * Bound here. However, this would be wrong, as we could
4907 * add an exception Proj to a new location then.
4908 * So, we must turn in into a tuple.
4914 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4915 switch (get_Proj_proj(proj)) {
4917 DBG_OPT_EXC_REM(proj);
4918 proj = get_Bound_mem(bound);
4920 case pn_Bound_X_except:
4921 DBG_OPT_EXC_REM(proj);
4922 proj = get_irg_bad(get_irn_irg(proj));
4926 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4928 case pn_Bound_X_regular:
4929 DBG_OPT_EXC_REM(proj);
4930 proj = new_r_Jmp(get_nodes_block(bound));
4937 } /* transform_node_Proj_Bound */
4940 * Does all optimizations on nodes that must be done on it's Proj's
4941 * because of creating new nodes.
4943 static ir_node *transform_node_Proj(ir_node *proj)
4945 ir_node *n = get_Proj_pred(proj);
4947 if (n->op->ops.transform_node_Proj)
4948 return n->op->ops.transform_node_Proj(proj);
4950 } /* transform_node_Proj */
4953 * Move Confirms down through Phi nodes.
4955 static ir_node *transform_node_Phi(ir_node *phi)
4958 ir_mode *mode = get_irn_mode(phi);
4960 if (mode_is_reference(mode)) {
4961 n = get_irn_arity(phi);
4963 /* Beware of Phi0 */
4965 ir_node *pred = get_irn_n(phi, 0);
4966 ir_node *bound, *new_Phi, *block, **in;
4969 if (! is_Confirm(pred))
4972 bound = get_Confirm_bound(pred);
4973 pnc = get_Confirm_cmp(pred);
4975 NEW_ARR_A(ir_node *, in, n);
4976 in[0] = get_Confirm_value(pred);
4978 for (i = 1; i < n; ++i) {
4979 pred = get_irn_n(phi, i);
4981 if (! is_Confirm(pred) ||
4982 get_Confirm_bound(pred) != bound ||
4983 get_Confirm_cmp(pred) != pnc)
4985 in[i] = get_Confirm_value(pred);
4987 /* move the Confirm nodes "behind" the Phi */
4988 block = get_irn_n(phi, -1);
4989 new_Phi = new_r_Phi(block, n, in, get_irn_mode(phi));
4990 return new_r_Confirm(block, new_Phi, bound, pnc);
4994 } /* transform_node_Phi */
4997 * Returns the operands of a commutative bin-op, if one operand is
4998 * a const, it is returned as the second one.
5000 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c)
5002 ir_node *op_a = get_binop_left(binop);
5003 ir_node *op_b = get_binop_right(binop);
5005 assert(is_op_commutative(get_irn_op(binop)));
5007 if (is_Const(op_a)) {
5014 } /* get_comm_Binop_Ops */
5017 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
5018 * Such pattern may arise in bitfield stores.
5020 * value c4 value c4 & c2
5021 * AND c3 AND c1 | c3
5028 * AND c1 ===> OR if (c1 | c2) == 0x111..11
5031 static ir_node *transform_node_Or_bf_store(ir_node *or)
5035 ir_node *and_l, *c3;
5036 ir_node *value, *c4;
5037 ir_node *new_and, *new_const, *block;
5038 ir_mode *mode = get_irn_mode(or);
5040 tarval *tv1, *tv2, *tv3, *tv4, *tv;
5043 get_comm_Binop_Ops(or, &and, &c1);
5044 if (!is_Const(c1) || !is_And(and))
5047 get_comm_Binop_Ops(and, &or_l, &c2);
5051 tv1 = get_Const_tarval(c1);
5052 tv2 = get_Const_tarval(c2);
5054 tv = tarval_or(tv1, tv2);
5055 if (tarval_is_all_one(tv)) {
5056 /* the AND does NOT clear a bit with isn't set by the OR */
5057 set_Or_left(or, or_l);
5058 set_Or_right(or, c1);
5060 /* check for more */
5067 get_comm_Binop_Ops(or_l, &and_l, &c3);
5068 if (!is_Const(c3) || !is_And(and_l))
5071 get_comm_Binop_Ops(and_l, &value, &c4);
5075 /* ok, found the pattern, check for conditions */
5076 assert(mode == get_irn_mode(and));
5077 assert(mode == get_irn_mode(or_l));
5078 assert(mode == get_irn_mode(and_l));
5080 tv3 = get_Const_tarval(c3);
5081 tv4 = get_Const_tarval(c4);
5083 tv = tarval_or(tv4, tv2);
5084 if (!tarval_is_all_one(tv)) {
5085 /* have at least one 0 at the same bit position */
5089 if (tv3 != tarval_andnot(tv3, tv4)) {
5090 /* bit in the or_mask is outside the and_mask */
5094 if (tv1 != tarval_andnot(tv1, tv2)) {
5095 /* bit in the or_mask is outside the and_mask */
5099 /* ok, all conditions met */
5100 block = get_irn_n(or, -1);
5102 new_and = new_r_And(block, value, new_Const(tarval_and(tv4, tv2)), mode);
5104 new_const = new_Const(tarval_or(tv3, tv1));
5106 set_Or_left(or, new_and);
5107 set_Or_right(or, new_const);
5109 /* check for more */
5111 } /* transform_node_Or_bf_store */
5114 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
5116 static ir_node *transform_node_Or_Rotl(ir_node *or)
5118 ir_mode *mode = get_irn_mode(or);
5119 ir_node *shl, *shr, *block;
5120 ir_node *irn, *x, *c1, *c2, *v, *sub, *n, *rotval;
5123 if (! mode_is_int(mode))
5126 shl = get_binop_left(or);
5127 shr = get_binop_right(or);
5136 } else if (!is_Shl(shl)) {
5138 } else if (!is_Shr(shr)) {
5141 x = get_Shl_left(shl);
5142 if (x != get_Shr_left(shr))
5145 c1 = get_Shl_right(shl);
5146 c2 = get_Shr_right(shr);
5147 if (is_Const(c1) && is_Const(c2)) {
5148 tv1 = get_Const_tarval(c1);
5149 if (! tarval_is_long(tv1))
5152 tv2 = get_Const_tarval(c2);
5153 if (! tarval_is_long(tv2))
5156 if (get_tarval_long(tv1) + get_tarval_long(tv2)
5157 != (int) get_mode_size_bits(mode))
5160 /* yet, condition met */
5161 block = get_nodes_block(or);
5163 n = new_r_Rotl(block, x, c1, mode);
5165 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
5172 rotval = sub; /* a Rot right is not supported, so use a rot left */
5173 } else if (is_Sub(c2)) {
5179 if (get_Sub_right(sub) != v)
5182 c1 = get_Sub_left(sub);
5186 tv1 = get_Const_tarval(c1);
5187 if (! tarval_is_long(tv1))
5190 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
5193 /* yet, condition met */
5194 block = get_nodes_block(or);
5196 n = new_r_Rotl(block, x, rotval, mode);
5198 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL);
5200 } /* transform_node_Or_Rotl */
5205 static ir_node *transform_node_Or(ir_node *n)
5207 ir_node *c, *oldn = n;
5208 ir_node *a = get_Or_left(n);
5209 ir_node *b = get_Or_right(n);
5212 if (is_Not(a) && is_Not(b)) {
5213 /* ~a | ~b = ~(a&b) */
5214 ir_node *block = get_nodes_block(n);
5216 mode = get_irn_mode(n);
5219 n = new_rd_And(get_irn_dbg_info(n), block, a, b, mode);
5220 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
5221 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
5225 /* we can evaluate 2 Projs of the same Cmp */
5226 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
5227 ir_node *pred_a = get_Proj_pred(a);
5228 ir_node *pred_b = get_Proj_pred(b);
5229 if (pred_a == pred_b) {
5230 dbg_info *dbgi = get_irn_dbg_info(n);
5231 pn_Cmp pn_a = get_Proj_proj(a);
5232 pn_Cmp pn_b = get_Proj_proj(b);
5233 /* yes, we can simply calculate with pncs */
5234 pn_Cmp new_pnc = pn_a | pn_b;
5236 return new_rd_Proj(dbgi, pred_a, mode_b, new_pnc);
5240 mode = get_irn_mode(n);
5241 HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode);
5243 n = transform_node_Or_bf_store(n);
5244 n = transform_node_Or_Rotl(n);
5248 n = transform_bitwise_distributive(n, transform_node_Or);
5251 } /* transform_node_Or */
5255 static ir_node *transform_node(ir_node *n);
5258 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
5260 * Should be moved to reassociation?
5262 static ir_node *transform_node_shift(ir_node *n)
5264 ir_node *left, *right;
5266 tarval *tv1, *tv2, *res;
5267 ir_node *in[2], *irn, *block;
5269 left = get_binop_left(n);
5271 /* different operations */
5272 if (get_irn_op(left) != get_irn_op(n))
5275 right = get_binop_right(n);
5276 tv1 = value_of(right);
5277 if (tv1 == tarval_bad)
5280 tv2 = value_of(get_binop_right(left));
5281 if (tv2 == tarval_bad)
5284 res = tarval_add(tv1, tv2);
5285 mode = get_irn_mode(n);
5287 /* beware: a simple replacement works only, if res < modulo shift */
5289 int modulo_shf = get_mode_modulo_shift(mode);
5290 if (modulo_shf > 0) {
5291 tarval *modulo = new_tarval_from_long(modulo_shf,
5292 get_tarval_mode(res));
5294 assert(modulo_shf >= (int) get_mode_size_bits(mode));
5296 /* shifting too much */
5297 if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) {
5299 ir_node *block = get_nodes_block(n);
5300 dbg_info *dbgi = get_irn_dbg_info(n);
5301 ir_mode *smode = get_irn_mode(right);
5302 ir_node *cnst = new_Const_long(smode, get_mode_size_bits(mode) - 1);
5303 return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode);
5306 return new_Const(get_mode_null(mode));
5310 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
5313 /* ok, we can replace it */
5314 block = get_nodes_block(n);
5316 in[0] = get_binop_left(left);
5317 in[1] = new_Const(res);
5319 irn = new_ir_node(NULL, get_Block_irg(block), block, get_irn_op(n), mode, 2, in);
5321 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5323 return transform_node(irn);
5324 } /* transform_node_shift */
5327 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
5329 * - and, or, xor instead of &
5330 * - Shl, Shr, Shrs, rotl instead of >>
5331 * (with a special case for Or/Xor + Shrs)
5333 static ir_node *transform_node_bitop_shift(ir_node *n)
5336 ir_node *right = get_binop_right(n);
5337 ir_mode *mode = get_irn_mode(n);
5338 ir_node *bitop_left;
5339 ir_node *bitop_right;
5350 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
5352 if (!is_Const(right))
5355 left = get_binop_left(n);
5356 op_left = get_irn_op(left);
5357 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
5360 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
5361 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
5362 /* TODO: test if sign bit is affectes */
5366 bitop_right = get_binop_right(left);
5367 if (!is_Const(bitop_right))
5370 bitop_left = get_binop_left(left);
5372 block = get_nodes_block(n);
5373 dbgi = get_irn_dbg_info(n);
5374 tv1 = get_Const_tarval(bitop_right);
5375 tv2 = get_Const_tarval(right);
5377 assert(get_tarval_mode(tv1) == mode);
5380 new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode);
5381 tv_shift = tarval_shl(tv1, tv2);
5382 } else if (is_Shr(n)) {
5383 new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode);
5384 tv_shift = tarval_shr(tv1, tv2);
5385 } else if (is_Shrs(n)) {
5386 new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode);
5387 tv_shift = tarval_shrs(tv1, tv2);
5390 new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode);
5391 tv_shift = tarval_rotl(tv1, tv2);
5394 assert(get_tarval_mode(tv_shift) == mode);
5395 new_const = new_Const(tv_shift);
5397 if (op_left == op_And) {
5398 new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode);
5399 } else if (op_left == op_Or) {
5400 new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode);
5402 assert(op_left == op_Eor);
5403 new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode);
5411 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5413 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5414 * (also with x >>s c1 when c1>=c2)
5416 static ir_node *transform_node_shl_shr(ir_node *n)
5419 ir_node *right = get_binop_right(n);
5434 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5436 if (!is_Const(right))
5439 left = get_binop_left(n);
5440 mode = get_irn_mode(n);
5441 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5442 ir_node *shr_right = get_binop_right(left);
5444 if (!is_Const(shr_right))
5447 x = get_binop_left(left);
5448 tv_shr = get_Const_tarval(shr_right);
5449 tv_shl = get_Const_tarval(right);
5451 if (is_Shrs(left)) {
5452 /* shrs variant only allowed if c1 >= c2 */
5453 if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge))
5456 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5459 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5461 tv_mask = tarval_shl(tv_mask, tv_shl);
5462 } else if (is_Shr(n) && is_Shl(left)) {
5463 ir_node *shl_right = get_Shl_right(left);
5465 if (!is_Const(shl_right))
5468 x = get_Shl_left(left);
5469 tv_shr = get_Const_tarval(right);
5470 tv_shl = get_Const_tarval(shl_right);
5472 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5473 tv_mask = tarval_shr(tv_mask, tv_shr);
5478 if (get_tarval_mode(tv_shl) != get_tarval_mode(tv_shr)) {
5479 tv_shl = tarval_convert_to(tv_shl, get_tarval_mode(tv_shr));
5482 assert(tv_mask != tarval_bad);
5483 assert(get_tarval_mode(tv_mask) == mode);
5485 block = get_nodes_block(n);
5486 dbgi = get_irn_dbg_info(n);
5488 pnc = tarval_cmp(tv_shl, tv_shr);
5489 if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) {
5490 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5491 new_const = new_Const(tv_shift);
5493 new_shift = new_rd_Shrs(dbgi, block, x, new_const, mode);
5495 new_shift = new_rd_Shr(dbgi, block, x, new_const, mode);
5498 assert(pnc == pn_Cmp_Gt);
5499 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5500 new_const = new_Const(tv_shift);
5501 new_shift = new_rd_Shl(dbgi, block, x, new_const, mode);
5504 new_const = new_Const(tv_mask);
5505 new_and = new_rd_And(dbgi, block, new_shift, new_const, mode);
5513 static ir_node *transform_node_Shr(ir_node *n)
5515 ir_node *c, *oldn = n;
5516 ir_node *left = get_Shr_left(n);
5517 ir_node *right = get_Shr_right(n);
5518 ir_mode *mode = get_irn_mode(n);
5520 HANDLE_BINOP_PHI((eval_func) tarval_shr, left, right, c, mode);
5521 n = transform_node_shift(n);
5524 n = transform_node_shl_shr(n);
5526 n = transform_node_bitop_shift(n);
5529 } /* transform_node_Shr */
5534 static ir_node *transform_node_Shrs(ir_node *n)
5536 ir_node *c, *oldn = n;
5537 ir_node *a = get_Shrs_left(n);
5538 ir_node *b = get_Shrs_right(n);
5539 ir_mode *mode = get_irn_mode(n);
5541 HANDLE_BINOP_PHI((eval_func) tarval_shrs, a, b, c, mode);
5542 n = transform_node_shift(n);
5545 n = transform_node_bitop_shift(n);
5548 } /* transform_node_Shrs */
5553 static ir_node *transform_node_Shl(ir_node *n)
5555 ir_node *c, *oldn = n;
5556 ir_node *a = get_Shl_left(n);
5557 ir_node *b = get_Shl_right(n);
5558 ir_mode *mode = get_irn_mode(n);
5560 HANDLE_BINOP_PHI((eval_func) tarval_shl, a, b, c, mode);
5561 n = transform_node_shift(n);
5564 n = transform_node_shl_shr(n);
5566 n = transform_node_bitop_shift(n);
5569 } /* transform_node_Shl */
5574 static ir_node *transform_node_Rotl(ir_node *n)
5576 ir_node *c, *oldn = n;
5577 ir_node *a = get_Rotl_left(n);
5578 ir_node *b = get_Rotl_right(n);
5579 ir_mode *mode = get_irn_mode(n);
5581 HANDLE_BINOP_PHI((eval_func) tarval_rotl, a, b, c, mode);
5582 n = transform_node_shift(n);
5585 n = transform_node_bitop_shift(n);
5588 } /* transform_node_Rotl */
5593 static ir_node *transform_node_Conv(ir_node *n)
5595 ir_node *c, *oldn = n;
5596 ir_mode *mode = get_irn_mode(n);
5597 ir_node *a = get_Conv_op(n);
5599 if (mode != mode_b && is_const_Phi(a)) {
5600 /* Do NOT optimize mode_b Conv's, this leads to remaining
5601 * Phib nodes later, because the conv_b_lower operation
5602 * is instantly reverted, when it tries to insert a Convb.
5604 c = apply_conv_on_phi(a, mode);
5606 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5611 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5612 ir_graph *irg = get_irn_irg(n);
5613 return new_r_Unknown(irg, mode);
5616 if (mode_is_reference(mode) &&
5617 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5619 ir_node *l = get_Add_left(a);
5620 ir_node *r = get_Add_right(a);
5621 dbg_info *dbgi = get_irn_dbg_info(a);
5622 ir_node *block = get_nodes_block(n);
5624 ir_node *lop = get_Conv_op(l);
5625 if (get_irn_mode(lop) == mode) {
5626 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5627 n = new_rd_Add(dbgi, block, lop, r, mode);
5632 ir_node *rop = get_Conv_op(r);
5633 if (get_irn_mode(rop) == mode) {
5634 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5635 n = new_rd_Add(dbgi, block, l, rop, mode);
5642 } /* transform_node_Conv */
5645 * Remove dead blocks and nodes in dead blocks
5646 * in keep alive list. We do not generate a new End node.
5648 static ir_node *transform_node_End(ir_node *n)
5650 int i, j, n_keepalives = get_End_n_keepalives(n);
5653 NEW_ARR_A(ir_node *, in, n_keepalives);
5655 for (i = j = 0; i < n_keepalives; ++i) {
5656 ir_node *ka = get_End_keepalive(n, i);
5658 if (! is_Block_dead(ka)) {
5662 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
5664 } else if (is_Bad(ka)) {
5665 /* no need to keep Bad */
5670 if (j != n_keepalives)
5671 set_End_keepalives(n, j, in);
5673 } /* transform_node_End */
5675 bool is_negated_value(ir_node *a, ir_node *b)
5677 if (is_Minus(a) && get_Minus_op(a) == b)
5679 if (is_Minus(b) && get_Minus_op(b) == a)
5681 if (is_Sub(a) && is_Sub(b)) {
5682 ir_node *a_left = get_Sub_left(a);
5683 ir_node *a_right = get_Sub_right(a);
5684 ir_node *b_left = get_Sub_left(b);
5685 ir_node *b_right = get_Sub_right(b);
5687 if (a_left == b_right && a_right == b_left)
5695 * Optimize a Mux into some simpler cases.
5697 static ir_node *transform_node_Mux(ir_node *n)
5699 ir_node *oldn = n, *sel = get_Mux_sel(n);
5700 ir_mode *mode = get_irn_mode(n);
5701 ir_node *t = get_Mux_true(n);
5702 ir_node *f = get_Mux_false(n);
5703 ir_graph *irg = current_ir_graph;
5706 ir_node* block = get_nodes_block(n);
5708 ir_node* c1 = get_Mux_sel(t);
5709 ir_node* t1 = get_Mux_true(t);
5710 ir_node* f1 = get_Mux_false(t);
5712 /* Mux(cond0, Mux(cond1, x, y), y) -> typical if (cond0 && cond1) x else y */
5713 ir_node* and_ = new_r_And(block, c0, c1, mode_b);
5714 ir_node* new_mux = new_r_Mux(block, and_, f1, t1, mode);
5719 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5720 } else if (f == t1) {
5721 /* Mux(cond0, Mux(cond1, x, y), x) */
5722 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5723 ir_node* and_ = new_r_And(block, c0, not_c1, mode_b);
5724 ir_node* new_mux = new_r_Mux(block, and_, t1, f1, mode);
5729 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5731 } else if (is_Mux(f)) {
5732 ir_node* block = get_nodes_block(n);
5734 ir_node* c1 = get_Mux_sel(f);
5735 ir_node* t1 = get_Mux_true(f);
5736 ir_node* f1 = get_Mux_false(f);
5738 /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */
5739 ir_node* or_ = new_r_Or(block, c0, c1, mode_b);
5740 ir_node* new_mux = new_r_Mux(block, or_, f1, t1, mode);
5745 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5746 } else if (t == f1) {
5747 /* Mux(cond0, x, Mux(cond1, y, x)) */
5748 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5749 ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b);
5750 ir_node* new_mux = new_r_Mux(block, or_, t1, f1, mode);
5755 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5759 /* first normalization step: move a possible zero to the false case */
5761 ir_node *cmp = get_Proj_pred(sel);
5764 if (is_Const(t) && is_Const_null(t)) {
5767 /* Mux(x, 0, y) => Mux(x, y, 0) */
5768 pn_Cmp pnc = get_Proj_proj(sel);
5769 sel = new_r_Proj(cmp, mode_b,
5770 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
5771 n = new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, t, f, mode);
5779 /* note: after normalization, false can only happen on default */
5780 if (mode == mode_b) {
5781 dbg_info *dbg = get_irn_dbg_info(n);
5782 ir_node *block = get_nodes_block(n);
5785 tarval *tv_t = get_Const_tarval(t);
5786 if (tv_t == tarval_b_true) {
5788 /* Muxb(sel, true, false) = sel */
5789 assert(get_Const_tarval(f) == tarval_b_false);
5790 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5793 /* Muxb(sel, true, x) = Or(sel, x) */
5794 n = new_rd_Or(dbg, block, sel, f, mode_b);
5795 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5799 } else if (is_Const(f)) {
5800 tarval *tv_f = get_Const_tarval(f);
5801 if (tv_f == tarval_b_true) {
5802 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5803 ir_node* not_sel = new_rd_Not(dbg, block, sel, mode_b);
5804 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5805 n = new_rd_Or(dbg, block, not_sel, t, mode_b);
5808 /* Muxb(sel, x, false) = And(sel, x) */
5809 assert(tv_f == tarval_b_false);
5810 n = new_rd_And(dbg, block, sel, t, mode_b);
5811 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5817 /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */
5818 if (is_Const(t) && is_Const(f) && mode_is_int(mode)
5819 && !is_irg_state(irg, IR_GRAPH_STATE_KEEP_MUX)) {
5820 tarval *a = get_Const_tarval(t);
5821 tarval *b = get_Const_tarval(f);
5824 if (tarval_is_one(a) && tarval_is_null(b)) {
5825 ir_node *block = get_nodes_block(n);
5826 ir_node *conv = new_r_Conv(block, sel, mode);
5828 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5830 } else if (tarval_is_null(a) && tarval_is_one(b)) {
5831 ir_node *block = get_nodes_block(n);
5832 ir_node *not_ = new_r_Not(block, sel, mode_b);
5833 ir_node *conv = new_r_Conv(block, not_, mode);
5835 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5839 /* TODO: it's not really clear if that helps in general or should be moved
5840 * to backend, especially with the MUX->Conv transformation above */
5841 if (tarval_cmp(a, b) & pn_Cmp_Gt) {
5842 diff = tarval_sub(a, b, NULL);
5845 diff = tarval_sub(b, a, NULL);
5849 if (diff == get_tarval_one(mode)) {
5850 dbg_info *dbg = get_irn_dbg_info(n);
5851 ir_node *block = get_nodes_block(n);
5852 ir_node *t = new_Const(tarval_sub(a, min, NULL));
5853 ir_node *f = new_Const(tarval_sub(b, min, NULL));
5854 n = new_rd_Mux(dbg, block, sel, f, t, mode);
5855 n = new_rd_Add(dbg, block, n, new_Const(min), mode);
5861 ir_node *cmp = get_Proj_pred(sel);
5862 long pn = get_Proj_proj(sel);
5865 * Note: normalization puts the constant on the right side,
5866 * so we check only one case.
5868 * Note further that these optimization work even for floating point
5869 * with NaN's because -NaN == NaN.
5870 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5874 ir_node *cmp_r = get_Cmp_right(cmp);
5875 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5876 ir_node *block = get_nodes_block(n);
5877 ir_node *cmp_l = get_Cmp_left(cmp);
5879 if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
5882 /* NaN's work fine with abs, so it is ok to remove Uo */
5883 long pnc = pn & ~pn_Cmp_Uo;
5885 if ( (cmp_l == t && (pnc == pn_Cmp_Ge || pnc == pn_Cmp_Gt))
5886 || (cmp_l == f && (pnc == pn_Cmp_Le || pnc == pn_Cmp_Lt)))
5888 /* Mux(a >/>= 0, a, -a) = Mux(a </<= 0, -a, a) ==> Abs(a) */
5889 n = new_rd_Abs(get_irn_dbg_info(n), block, cmp_l, mode);
5890 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5892 } else if ((cmp_l == t && (pnc == pn_Cmp_Le || pnc == pn_Cmp_Lt))
5893 || (cmp_l == f && (pnc == pn_Cmp_Ge || pnc == pn_Cmp_Gt)))
5895 /* Mux(a </<= 0, a, -a) = Mux(a >/>= 0, -a, a) ==> -Abs(a) */
5896 n = new_rd_Abs(get_irn_dbg_info(n), block, cmp_l, mode);
5897 n = new_rd_Minus(get_irn_dbg_info(n), block, n, mode);
5898 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5903 if (mode_is_int(mode)) {
5905 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
5906 /* Mux((a & b) != 0, c, 0) */
5907 ir_node *and_r = get_And_right(cmp_l);
5910 if (and_r == t && f == cmp_r) {
5911 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5912 if (pn == pn_Cmp_Lg) {
5913 /* Mux((a & 2^C) != 0, 2^C, 0) */
5915 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5917 /* Mux((a & 2^C) == 0, 2^C, 0) */
5918 n = new_rd_Eor(get_irn_dbg_info(n),
5919 block, cmp_l, t, mode);
5920 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5925 if (is_Shl(and_r)) {
5926 ir_node *shl_l = get_Shl_left(and_r);
5927 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5928 if (and_r == t && f == cmp_r) {
5929 if (pn == pn_Cmp_Lg) {
5930 /* (a & (1 << n)) != 0, (1 << n), 0) */
5932 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5934 /* (a & (1 << n)) == 0, (1 << n), 0) */
5935 n = new_rd_Eor(get_irn_dbg_info(n),
5936 block, cmp_l, t, mode);
5937 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5943 and_l = get_And_left(cmp_l);
5944 if (is_Shl(and_l)) {
5945 ir_node *shl_l = get_Shl_left(and_l);
5946 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5947 if (and_l == t && f == cmp_r) {
5948 if (pn == pn_Cmp_Lg) {
5949 /* ((1 << n) & a) != 0, (1 << n), 0) */
5951 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5953 /* ((1 << n) & a) == 0, (1 << n), 0) */
5954 n = new_rd_Eor(get_irn_dbg_info(n),
5955 block, cmp_l, t, mode);
5956 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5969 } /* transform_node_Mux */
5972 * optimize Sync nodes that have other syncs as input we simply add the inputs
5973 * of the other sync to our own inputs
5975 static ir_node *transform_node_Sync(ir_node *n)
5977 int arity = get_Sync_n_preds(n);
5980 for (i = 0; i < arity;) {
5981 ir_node *pred = get_Sync_pred(n, i);
5985 if (!is_Sync(pred)) {
5993 pred_arity = get_Sync_n_preds(pred);
5994 for (j = 0; j < pred_arity; ++j) {
5995 ir_node *pred_pred = get_Sync_pred(pred, j);
6000 add_irn_n(n, pred_pred);
6004 if (get_Sync_pred(n, k) == pred_pred) break;
6009 /* rehash the sync node */
6013 } /* transform_node_Sync */
6016 * optimize a trampoline Call into a direct Call
6018 static ir_node *transform_node_Call(ir_node *call)
6020 ir_node *callee = get_Call_ptr(call);
6021 ir_node *adr, *mem, *res, *bl, **in;
6022 ir_type *ctp, *mtp, *tp;
6026 int i, n_res, n_param;
6029 if (! is_Proj(callee))
6031 callee = get_Proj_pred(callee);
6032 if (! is_Builtin(callee))
6034 if (get_Builtin_kind(callee) != ir_bk_inner_trampoline)
6037 mem = get_Call_mem(call);
6039 if (skip_Proj(mem) == callee) {
6040 /* memory is routed to the trampoline, skip */
6041 mem = get_Builtin_mem(callee);
6044 /* build a new call type */
6045 mtp = get_Call_type(call);
6046 tdb = get_type_dbg_info(mtp);
6048 n_res = get_method_n_ress(mtp);
6049 n_param = get_method_n_params(mtp);
6050 ctp = new_d_type_method(n_param + 1, n_res, tdb);
6052 for (i = 0; i < n_res; ++i)
6053 set_method_res_type(ctp, i, get_method_res_type(mtp, i));
6055 NEW_ARR_A(ir_node *, in, n_param + 1);
6057 /* FIXME: we don't need a new pointer type in every step */
6058 irg = get_irn_irg(call);
6059 tp = get_irg_frame_type(irg);
6060 tp = new_type_pointer(tp);
6061 set_method_param_type(ctp, 0, tp);
6063 in[0] = get_Builtin_param(callee, 2);
6064 for (i = 0; i < n_param; ++i) {
6065 set_method_param_type(ctp, i + 1, get_method_param_type(mtp, i));
6066 in[i + 1] = get_Call_param(call, i);
6068 var = get_method_variadicity(mtp);
6069 set_method_variadicity(ctp, var);
6070 if (var == variadicity_variadic) {
6071 set_method_first_variadic_param_index(ctp, get_method_first_variadic_param_index(mtp) + 1);
6073 /* When we resolve a trampoline, the function must be called by a this-call */
6074 set_method_calling_convention(ctp, get_method_calling_convention(mtp) | cc_this_call);
6075 set_method_additional_properties(ctp, get_method_additional_properties(mtp));
6077 adr = get_Builtin_param(callee, 1);
6079 db = get_irn_dbg_info(call);
6080 bl = get_nodes_block(call);
6082 res = new_rd_Call(db, bl, mem, adr, n_param + 1, in, ctp);
6083 if (get_irn_pinned(call) == op_pin_state_floats)
6084 set_irn_pinned(res, op_pin_state_floats);
6086 } /* transform_node_Call */
6089 * Tries several [inplace] [optimizing] transformations and returns an
6090 * equivalent node. The difference to equivalent_node() is that these
6091 * transformations _do_ generate new nodes, and thus the old node must
6092 * not be freed even if the equivalent node isn't the old one.
6094 static ir_node *transform_node(ir_node *n)
6099 * Transform_node is the only "optimizing transformation" that might
6100 * return a node with a different opcode. We iterate HERE until fixpoint
6101 * to get the final result.
6105 if (n->op->ops.transform_node != NULL)
6106 n = n->op->ops.transform_node(n);
6107 } while (oldn != n);
6110 } /* transform_node */
6113 * Sets the default transform node operation for an ir_op_ops.
6115 * @param code the opcode for the default operation
6116 * @param ops the operations initialized
6121 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
6125 ops->transform_node = transform_node_##a; \
6127 #define CASE_PROJ(a) \
6129 ops->transform_node_Proj = transform_node_Proj_##a; \
6131 #define CASE_PROJ_EX(a) \
6133 ops->transform_node = transform_node_##a; \
6134 ops->transform_node_Proj = transform_node_Proj_##a; \
6143 CASE_PROJ_EX(DivMod);
6178 } /* firm_set_default_transform_node */
6181 /* **************** Common Subexpression Elimination **************** */
6183 /** The size of the hash table used, should estimate the number of nodes
6185 #define N_IR_NODES 512
6187 /** Compares the attributes of two Const nodes. */
6188 static int node_cmp_attr_Const(ir_node *a, ir_node *b)
6190 return (get_Const_tarval(a) != get_Const_tarval(b))
6191 || (get_Const_type(a) != get_Const_type(b));
6192 } /* node_cmp_attr_Const */
6194 /** Compares the attributes of two Proj nodes. */
6195 static int node_cmp_attr_Proj(ir_node *a, ir_node *b)
6197 return a->attr.proj != b->attr.proj;
6198 } /* node_cmp_attr_Proj */
6200 /** Compares the attributes of two Alloc nodes. */
6201 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b)
6203 const alloc_attr *pa = &a->attr.alloc;
6204 const alloc_attr *pb = &b->attr.alloc;
6205 return (pa->where != pb->where) || (pa->type != pb->type);
6206 } /* node_cmp_attr_Alloc */
6208 /** Compares the attributes of two Free nodes. */
6209 static int node_cmp_attr_Free(ir_node *a, ir_node *b)
6211 const free_attr *pa = &a->attr.free;
6212 const free_attr *pb = &b->attr.free;
6213 return (pa->where != pb->where) || (pa->type != pb->type);
6214 } /* node_cmp_attr_Free */
6216 /** Compares the attributes of two SymConst nodes. */
6217 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b)
6219 const symconst_attr *pa = &a->attr.symc;
6220 const symconst_attr *pb = &b->attr.symc;
6221 return (pa->kind != pb->kind)
6222 || (pa->sym.type_p != pb->sym.type_p)
6223 || (pa->tp != pb->tp);
6224 } /* node_cmp_attr_SymConst */
6226 /** Compares the attributes of two Call nodes. */
6227 static int node_cmp_attr_Call(ir_node *a, ir_node *b)
6229 const call_attr *pa = &a->attr.call;
6230 const call_attr *pb = &b->attr.call;
6231 return (pa->type != pb->type)
6232 || (pa->tail_call != pb->tail_call);
6233 } /* node_cmp_attr_Call */
6235 /** Compares the attributes of two Sel nodes. */
6236 static int node_cmp_attr_Sel(ir_node *a, ir_node *b)
6238 const ir_entity *a_ent = get_Sel_entity(a);
6239 const ir_entity *b_ent = get_Sel_entity(b);
6240 return a_ent != b_ent;
6241 } /* node_cmp_attr_Sel */
6243 /** Compares the attributes of two Phi nodes. */
6244 static int node_cmp_attr_Phi(ir_node *a, ir_node *b)
6246 /* we can only enter this function if both nodes have the same number of inputs,
6247 hence it is enough to check if one of them is a Phi0 */
6249 /* check the Phi0 pos attribute */
6250 return a->attr.phi.u.pos != b->attr.phi.u.pos;
6253 } /* node_cmp_attr_Phi */
6255 /** Compares the attributes of two Conv nodes. */
6256 static int node_cmp_attr_Conv(ir_node *a, ir_node *b)
6258 return get_Conv_strict(a) != get_Conv_strict(b);
6259 } /* node_cmp_attr_Conv */
6261 /** Compares the attributes of two Cast nodes. */
6262 static int node_cmp_attr_Cast(ir_node *a, ir_node *b)
6264 return get_Cast_type(a) != get_Cast_type(b);
6265 } /* node_cmp_attr_Cast */
6267 /** Compares the attributes of two Load nodes. */
6268 static int node_cmp_attr_Load(ir_node *a, ir_node *b)
6270 if (get_Load_volatility(a) == volatility_is_volatile ||
6271 get_Load_volatility(b) == volatility_is_volatile)
6272 /* NEVER do CSE on volatile Loads */
6274 /* do not CSE Loads with different alignment. Be conservative. */
6275 if (get_Load_align(a) != get_Load_align(b))
6278 return get_Load_mode(a) != get_Load_mode(b);
6279 } /* node_cmp_attr_Load */
6281 /** Compares the attributes of two Store nodes. */
6282 static int node_cmp_attr_Store(ir_node *a, ir_node *b)
6284 /* do not CSE Stores with different alignment. Be conservative. */
6285 if (get_Store_align(a) != get_Store_align(b))
6288 /* NEVER do CSE on volatile Stores */
6289 return (get_Store_volatility(a) == volatility_is_volatile ||
6290 get_Store_volatility(b) == volatility_is_volatile);
6291 } /* node_cmp_attr_Store */
6293 /** Compares two exception attributes */
6294 static int node_cmp_exception(ir_node *a, ir_node *b)
6296 const except_attr *ea = &a->attr.except;
6297 const except_attr *eb = &b->attr.except;
6299 return ea->pin_state != eb->pin_state;
6302 #define node_cmp_attr_Bound node_cmp_exception
6304 /** Compares the attributes of two Div nodes. */
6305 static int node_cmp_attr_Div(ir_node *a, ir_node *b)
6307 const divmod_attr *ma = &a->attr.divmod;
6308 const divmod_attr *mb = &b->attr.divmod;
6309 return ma->exc.pin_state != mb->exc.pin_state ||
6310 ma->resmode != mb->resmode ||
6311 ma->no_remainder != mb->no_remainder;
6312 } /* node_cmp_attr_Div */
6314 /** Compares the attributes of two DivMod nodes. */
6315 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b)
6317 const divmod_attr *ma = &a->attr.divmod;
6318 const divmod_attr *mb = &b->attr.divmod;
6319 return ma->exc.pin_state != mb->exc.pin_state ||
6320 ma->resmode != mb->resmode;
6321 } /* node_cmp_attr_DivMod */
6323 /** Compares the attributes of two Mod nodes. */
6324 static int node_cmp_attr_Mod(ir_node *a, ir_node *b)
6326 return node_cmp_attr_DivMod(a, b);
6327 } /* node_cmp_attr_Mod */
6329 /** Compares the attributes of two Quot nodes. */
6330 static int node_cmp_attr_Quot(ir_node *a, ir_node *b)
6332 return node_cmp_attr_DivMod(a, b);
6333 } /* node_cmp_attr_Quot */
6335 /** Compares the attributes of two Confirm nodes. */
6336 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b)
6338 /* no need to compare the bound, as this is a input */
6339 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
6340 } /* node_cmp_attr_Confirm */
6342 /** Compares the attributes of two Builtin nodes. */
6343 static int node_cmp_attr_Builtin(ir_node *a, ir_node *b)
6345 /* no need to compare the type, equal kind means equal type */
6346 return get_Builtin_kind(a) != get_Builtin_kind(b);
6347 } /* node_cmp_attr_Builtin */
6349 /** Compares the attributes of two ASM nodes. */
6350 static int node_cmp_attr_ASM(ir_node *a, ir_node *b)
6353 const ir_asm_constraint *ca;
6354 const ir_asm_constraint *cb;
6357 if (get_ASM_text(a) != get_ASM_text(b))
6360 /* Should we really check the constraints here? Should be better, but is strange. */
6361 n = get_ASM_n_input_constraints(a);
6362 if (n != get_ASM_n_input_constraints(b))
6365 ca = get_ASM_input_constraints(a);
6366 cb = get_ASM_input_constraints(b);
6367 for (i = 0; i < n; ++i) {
6368 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
6372 n = get_ASM_n_output_constraints(a);
6373 if (n != get_ASM_n_output_constraints(b))
6376 ca = get_ASM_output_constraints(a);
6377 cb = get_ASM_output_constraints(b);
6378 for (i = 0; i < n; ++i) {
6379 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
6383 n = get_ASM_n_clobbers(a);
6384 if (n != get_ASM_n_clobbers(b))
6387 cla = get_ASM_clobbers(a);
6388 clb = get_ASM_clobbers(b);
6389 for (i = 0; i < n; ++i) {
6390 if (cla[i] != clb[i])
6394 } /* node_cmp_attr_ASM */
6396 /** Compares the inexistent attributes of two Dummy nodes. */
6397 static int node_cmp_attr_Dummy(ir_node *a, ir_node *b)
6405 * Set the default node attribute compare operation for an ir_op_ops.
6407 * @param code the opcode for the default operation
6408 * @param ops the operations initialized
6413 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
6417 ops->node_cmp_attr = node_cmp_attr_##a; \
6450 } /* firm_set_default_node_cmp_attr */
6453 * Compare function for two nodes in the value table. Gets two
6454 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6456 int identities_cmp(const void *elt, const void *key)
6458 ir_node *a = (ir_node *)elt;
6459 ir_node *b = (ir_node *)key;
6462 if (a == b) return 0;
6464 if ((get_irn_op(a) != get_irn_op(b)) ||
6465 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6467 /* compare if a's in and b's in are of equal length */
6468 irn_arity_a = get_irn_arity(a);
6469 if (irn_arity_a != get_irn_arity(b))
6472 if (get_irn_pinned(a) == op_pin_state_pinned) {
6473 /* for pinned nodes, the block inputs must be equal */
6474 if (get_irn_n(a, -1) != get_irn_n(b, -1))
6476 } else if (! get_opt_global_cse()) {
6477 /* for block-local CSE both nodes must be in the same MacroBlock */
6478 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
6482 /* compare a->in[0..ins] with b->in[0..ins] */
6483 for (i = 0; i < irn_arity_a; ++i) {
6484 ir_node *pred_a = get_irn_n(a, i);
6485 ir_node *pred_b = get_irn_n(b, i);
6486 if (pred_a != pred_b) {
6487 /* if both predecessors are CSE neutral they might be different */
6488 if (!is_irn_cse_neutral(pred_a) || !is_irn_cse_neutral(pred_b))
6494 * here, we already now that the nodes are identical except their
6497 if (a->op->ops.node_cmp_attr)
6498 return a->op->ops.node_cmp_attr(a, b);
6501 } /* identities_cmp */
6504 * Calculate a hash value of a node.
6506 * @param node The IR-node
6508 unsigned ir_node_hash(const ir_node *node)
6510 return node->op->ops.hash(node);
6511 } /* ir_node_hash */
6514 void new_identities(ir_graph *irg)
6516 if (irg->value_table != NULL)
6517 del_pset(irg->value_table);
6518 irg->value_table = new_pset(identities_cmp, N_IR_NODES);
6519 } /* new_identities */
6521 void del_identities(ir_graph *irg)
6523 if (irg->value_table != NULL)
6524 del_pset(irg->value_table);
6525 } /* del_identities */
6527 /* Normalize a node by putting constants (and operands with larger
6528 * node index) on the right (operator side). */
6529 void ir_normalize_node(ir_node *n)
6531 if (is_op_commutative(get_irn_op(n))) {
6532 ir_node *l = get_binop_left(n);
6533 ir_node *r = get_binop_right(n);
6535 /* For commutative operators perform a OP b == b OP a but keep
6536 * constants on the RIGHT side. This helps greatly in some
6537 * optimizations. Moreover we use the idx number to make the form
6539 if (!operands_are_normalized(l, r)) {
6540 set_binop_left(n, r);
6541 set_binop_right(n, l);
6545 } /* ir_normalize_node */
6548 * Update the nodes after a match in the value table. If both nodes have
6549 * the same MacroBlock but different Blocks, we must ensure that the node
6550 * with the dominating Block (the node that is near to the MacroBlock header
6551 * is stored in the table.
6552 * Because a MacroBlock has only one "non-exception" flow, we don't need
6553 * dominance info here: We known, that one block must dominate the other and
6554 * following the only block input will allow to find it.
6556 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node)
6558 ir_node *known_blk, *new_block, *block, *mbh;
6560 if (get_opt_global_cse()) {
6561 /* Block inputs are meaning less */
6564 known_blk = get_irn_n(known_irn, -1);
6565 new_block = get_irn_n(new_ir_node, -1);
6566 if (known_blk == new_block) {
6567 /* already in the same block */
6571 * We expect the typical case when we built the graph. In that case, the
6572 * known_irn is already the upper one, so checking this should be faster.
6575 mbh = get_Block_MacroBlock(new_block);
6577 if (block == known_blk) {
6578 /* ok, we have found it: known_block dominates new_block as expected */
6583 * We have reached the MacroBlock header NOT founding
6584 * the known_block. new_block must dominate known_block.
6587 set_irn_n(known_irn, -1, new_block);
6590 assert(get_Block_n_cfgpreds(block) == 1);
6591 block = get_Block_cfgpred_block(block, 0);
6593 } /* update_value_table */
6596 * Return the canonical node computing the same value as n.
6597 * Looks up the node in a hash table, enters it in the table
6598 * if it isn't there yet.
6600 * @param n the node to look up
6602 * @return a node that computes the same value as n or n if no such
6603 * node could be found
6605 ir_node *identify_remember(ir_node *n)
6607 ir_graph *irg = get_irn_irg(n);
6608 pset *value_table = irg->value_table;
6611 if (value_table == NULL)
6614 ir_normalize_node(n);
6615 /* lookup or insert in hash table with given hash key. */
6616 nn = pset_insert(value_table, n, ir_node_hash(n));
6619 update_known_irn(nn, n);
6621 /* n is reachable again */
6622 edges_node_revival(nn, get_irn_irg(nn));
6626 } /* identify_remember */
6629 * During construction we set the op_pin_state_pinned flag in the graph right when the
6630 * optimization is performed. The flag turning on procedure global cse could
6631 * be changed between two allocations. This way we are safe.
6633 * @param n The node to lookup
6635 static inline ir_node *identify_cons(ir_node *n)
6639 n = identify_remember(n);
6640 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n)) {
6641 ir_graph *irg = get_irn_irg(n);
6642 set_irg_pinned(irg, op_pin_state_floats);
6645 } /* identify_cons */
6647 /* Add a node to the identities value table. */
6648 void add_identities(ir_node *node)
6655 identify_remember(node);
6658 /* Visit each node in the value table of a graph. */
6659 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env)
6662 ir_graph *rem = current_ir_graph;
6664 current_ir_graph = irg;
6665 foreach_pset(irg->value_table, node) {
6668 current_ir_graph = rem;
6669 } /* visit_all_identities */
6672 * Garbage in, garbage out. If a node has a dead input, i.e., the
6673 * Bad node is input to the node, return the Bad node.
6675 static ir_node *gigo(ir_node *node)
6678 ir_op *op = get_irn_op(node);
6680 /* remove garbage blocks by looking at control flow that leaves the block
6681 and replacing the control flow by Bad. */
6682 if (get_irn_mode(node) == mode_X) {
6683 ir_node *block = get_nodes_block(skip_Proj(node));
6685 /* Don't optimize nodes in immature blocks. */
6686 if (!get_Block_matured(block))
6688 /* Don't optimize End, may have Bads. */
6689 if (op == op_End) return node;
6691 if (is_Block(block)) {
6692 if (is_Block_dead(block)) {
6693 /* control flow from dead block is dead */
6697 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
6698 if (!is_Bad(get_irn_n(block, i)))
6702 ir_graph *irg = get_irn_irg(block);
6703 /* the start block is never dead */
6704 if (block != get_irg_start_block(irg)
6705 && block != get_irg_end_block(irg)) {
6707 * Do NOT kill control flow without setting
6708 * the block to dead of bad things can happen:
6709 * We get a Block that is not reachable be irg_block_walk()
6710 * but can be found by irg_walk()!
6712 set_Block_dead(block);
6719 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6720 blocks predecessors is dead. */
6721 if (op != op_Block && op != op_Phi && op != op_Tuple) {
6722 irn_arity = get_irn_arity(node);
6725 * Beware: we can only read the block of a non-floating node.
6727 if (is_irn_pinned_in_irg(node) &&
6728 is_Block_dead(get_nodes_block(skip_Proj(node))))
6731 for (i = 0; i < irn_arity; i++) {
6732 ir_node *pred = get_irn_n(node, i);
6737 /* Propagating Unknowns here seems to be a bad idea, because
6738 sometimes we need a node as a input and did not want that
6740 However, it might be useful to move this into a later phase
6741 (if you think that optimizing such code is useful). */
6742 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
6743 return new_Unknown(get_irn_mode(node));
6748 /* With this code we violate the agreement that local_optimize
6749 only leaves Bads in Block, Phi and Tuple nodes. */
6750 /* If Block has only Bads as predecessors it's garbage. */
6751 /* If Phi has only Bads as predecessors it's garbage. */
6752 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
6753 irn_arity = get_irn_arity(node);
6754 for (i = 0; i < irn_arity; i++) {
6755 if (!is_Bad(get_irn_n(node, i))) break;
6757 if (i == irn_arity) node = new_Bad();
6764 * These optimizations deallocate nodes from the obstack.
6765 * It can only be called if it is guaranteed that no other nodes
6766 * reference this one, i.e., right after construction of a node.
6768 * @param n The node to optimize
6770 ir_node *optimize_node(ir_node *n)
6774 ir_graph *irg = get_irn_irg(n);
6775 ir_opcode iro = get_irn_opcode(n);
6777 /* Always optimize Phi nodes: part of the construction. */
6778 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6780 /* constant expression evaluation / constant folding */
6781 if (get_opt_constant_folding()) {
6782 /* neither constants nor Tuple values can be evaluated */
6783 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6784 unsigned fp_model = get_irg_fp_model(irg);
6785 int old_fp_mode = tarval_fp_ops_enabled();
6787 tarval_enable_fp_ops(! (fp_model & fp_no_float_fold));
6789 /* try to evaluate */
6790 tv = computed_value(n);
6791 if (tv != tarval_bad) {
6793 ir_type *old_tp = get_irn_type(n);
6794 int i, arity = get_irn_arity(n);
6798 * Try to recover the type of the new expression.
6800 for (i = 0; i < arity && !old_tp; ++i)
6801 old_tp = get_irn_type(get_irn_n(n, i));
6804 * we MUST copy the node here temporary, because it's still needed
6805 * for DBG_OPT_CSTEVAL
6807 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6808 oldn = alloca(node_size);
6810 memcpy(oldn, n, node_size);
6811 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6813 /* ARG, copy the in array, we need it for statistics */
6814 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6816 /* note the inplace edges module */
6817 edges_node_deleted(n, irg);
6819 /* evaluation was successful -- replace the node. */
6820 irg_kill_node(irg, n);
6823 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6824 set_Const_type(nw, old_tp);
6825 DBG_OPT_CSTEVAL(oldn, nw);
6826 tarval_enable_fp_ops(old_fp_mode);
6829 tarval_enable_fp_ops(old_fp_mode);
6833 /* remove unnecessary nodes */
6834 if (get_opt_algebraic_simplification() ||
6835 (iro == iro_Phi) || /* always optimize these nodes. */
6837 (iro == iro_Proj) ||
6838 (iro == iro_Block) ) /* Flags tested local. */
6839 n = equivalent_node(n);
6841 /* Common Subexpression Elimination.
6843 * Checks whether n is already available.
6844 * The block input is used to distinguish different subexpressions. Right
6845 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6846 * subexpressions within a block.
6849 n = identify_cons(n);
6852 edges_node_deleted(oldn, irg);
6854 /* We found an existing, better node, so we can deallocate the old node. */
6855 irg_kill_node(irg, oldn);
6859 /* Some more constant expression evaluation that does not allow to
6861 iro = get_irn_opcode(n);
6862 if (get_opt_algebraic_simplification() ||
6863 (iro == iro_Cond) ||
6864 (iro == iro_Proj)) /* Flags tested local. */
6865 n = transform_node(n);
6867 /* Remove nodes with dead (Bad) input.
6868 Run always for transformation induced Bads. */
6871 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6872 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6874 n = identify_remember(o);
6880 } /* optimize_node */
6884 * These optimizations never deallocate nodes (in place). This can cause dead
6885 * nodes lying on the obstack. Remove these by a dead node elimination,
6886 * i.e., a copying garbage collection.
6888 ir_node *optimize_in_place_2(ir_node *n)
6892 ir_opcode iro = get_irn_opcode(n);
6894 if (!get_opt_optimize() && !is_Phi(n)) return n;
6896 /* constant expression evaluation / constant folding */
6897 if (get_opt_constant_folding()) {
6898 /* neither constants nor Tuple values can be evaluated */
6899 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6900 ir_graph *irg = get_irn_irg(n);
6901 unsigned fp_model = get_irg_fp_model(irg);
6902 int old_fp_mode = tarval_fp_ops_enabled();
6904 tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6905 /* try to evaluate */
6906 tv = computed_value(n);
6907 if (tv != tarval_bad) {
6908 /* evaluation was successful -- replace the node. */
6909 ir_type *old_tp = get_irn_type(n);
6910 int i, arity = get_irn_arity(n);
6913 * Try to recover the type of the new expression.
6915 for (i = 0; i < arity && !old_tp; ++i)
6916 old_tp = get_irn_type(get_irn_n(n, i));
6920 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6921 set_Const_type(n, old_tp);
6923 DBG_OPT_CSTEVAL(oldn, n);
6924 tarval_enable_fp_ops(old_fp_mode);
6927 tarval_enable_fp_ops(old_fp_mode);
6931 /* remove unnecessary nodes */
6932 if (get_opt_constant_folding() ||
6933 (iro == iro_Phi) || /* always optimize these nodes. */
6934 (iro == iro_Id) || /* ... */
6935 (iro == iro_Proj) || /* ... */
6936 (iro == iro_Block) ) /* Flags tested local. */
6937 n = equivalent_node(n);
6939 /** common subexpression elimination **/
6940 /* Checks whether n is already available. */
6941 /* The block input is used to distinguish different subexpressions. Right
6942 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6943 subexpressions within a block. */
6944 if (get_opt_cse()) {
6946 n = identify_remember(o);
6951 /* Some more constant expression evaluation. */
6952 iro = get_irn_opcode(n);
6953 if (get_opt_constant_folding() ||
6954 (iro == iro_Cond) ||
6955 (iro == iro_Proj)) /* Flags tested local. */
6956 n = transform_node(n);
6958 /* Remove nodes with dead (Bad) input.
6959 Run always for transformation induced Bads. */
6962 /* Now we can verify the node, as it has no dead inputs any more. */
6965 /* Now we have a legal, useful node. Enter it in hash table for cse.
6966 Blocks should be unique anyways. (Except the successor of start:
6967 is cse with the start block!) */
6968 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6970 n = identify_remember(o);
6976 } /* optimize_in_place_2 */
6979 * Wrapper for external use, set proper status bits after optimization.
6981 ir_node *optimize_in_place(ir_node *n)
6983 ir_graph *irg = get_irn_irg(n);
6984 /* Handle graph state */
6985 assert(get_irg_phase_state(irg) != phase_building);
6987 if (get_opt_global_cse())
6988 set_irg_pinned(irg, op_pin_state_floats);
6989 if (get_irg_outs_state(irg) == outs_consistent)
6990 set_irg_outs_inconsistent(irg);
6992 /* FIXME: Maybe we could also test whether optimizing the node can
6993 change the control graph. */
6994 set_irg_doms_inconsistent(irg);
6995 return optimize_in_place_2(n);
6996 } /* optimize_in_place */
6999 * Calculate a hash value of a Const node.
7001 static unsigned hash_Const(const ir_node *node)
7005 /* special value for const, as they only differ in their tarval. */
7006 h = HASH_PTR(node->attr.con.tarval);
7012 * Calculate a hash value of a SymConst node.
7014 static unsigned hash_SymConst(const ir_node *node)
7018 /* all others are pointers */
7019 h = HASH_PTR(node->attr.symc.sym.type_p);
7022 } /* hash_SymConst */
7025 * Set the default hash operation in an ir_op_ops.
7027 * @param code the opcode for the default operation
7028 * @param ops the operations initialized
7033 static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops)
7037 ops->hash = hash_##a; \
7040 /* hash function already set */
7041 if (ops->hash != NULL)
7048 /* use input/mode default hash if no function was given */
7049 ops->hash = firm_default_hash;
7057 * Sets the default operation for an ir_ops.
7059 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops)
7061 ops = firm_set_default_hash(code, ops);
7062 ops = firm_set_default_computed_value(code, ops);
7063 ops = firm_set_default_equivalent_node(code, ops);
7064 ops = firm_set_default_transform_node(code, ops);
7065 ops = firm_set_default_node_cmp_attr(code, ops);
7066 ops = firm_set_default_get_type(code, ops);
7067 ops = firm_set_default_get_type_attr(code, ops);
7068 ops = firm_set_default_get_entity_attr(code, ops);
7071 } /* firm_set_default_operations */