3 * File name: ir/ir/iropt.c
4 * Purpose: iropt --- optimizations intertwined with IR construction.
5 * Author: Christian Schaefer
6 * Modified by: Goetz Lindenmaier, Michael Beck
9 * Copyright: (c) 1998-2005 Universität Karlsruhe
10 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
27 # include "irnode_t.h"
28 # include "irgraph_t.h"
29 # include "iredges_t.h"
30 # include "irmode_t.h"
32 # include "ircons_t.h"
36 # include "dbginfo_t.h"
37 # include "iropt_dbg.h"
38 # include "irflag_t.h"
43 # include "opt_polymorphy.h"
44 # include "opt_confirms.h"
46 /* Make types visible to allow most efficient access */
47 # include "entity_t.h"
50 * return the value of a Constant
52 static tarval *computed_value_Const(ir_node *n)
54 return get_Const_tarval(n);
58 * return the value of a 'sizeof' SymConst
60 static tarval *computed_value_SymConst(ir_node *n)
62 if ((get_SymConst_kind(n) == symconst_size) &&
63 (get_type_state(get_SymConst_type(n))) == layout_fixed)
64 return new_tarval_from_long(get_type_size_bytes(get_SymConst_type(n)), get_irn_mode(n));
69 * return the value of an Add
71 static tarval *computed_value_Add(ir_node *n)
73 ir_node *a = get_Add_left(n);
74 ir_node *b = get_Add_right(n);
76 tarval *ta = value_of(a);
77 tarval *tb = value_of(b);
79 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b)))
80 return tarval_add(ta, tb);
86 * return the value of a Sub
89 static tarval *computed_value_Sub(ir_node *n)
91 ir_node *a = get_Sub_left(n);
92 ir_node *b = get_Sub_right(n);
97 if (a == b && !is_Bad(a))
98 return get_mode_null(get_irn_mode(n));
103 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b)))
104 return tarval_sub(ta, tb);
110 * return the value of an unary Minus
112 static tarval *computed_value_Minus(ir_node *n)
114 ir_node *a = get_Minus_op(n);
115 tarval *ta = value_of(a);
117 if ((ta != tarval_bad) && mode_is_signed(get_irn_mode(a)))
118 return tarval_neg(ta);
124 * return the value of a Mul
126 static tarval *computed_value_Mul(ir_node *n)
128 ir_node *a = get_Mul_left(n);
129 ir_node *b = get_Mul_right(n);
131 tarval *ta = value_of(a);
132 tarval *tb = value_of(b);
134 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
135 return tarval_mul(ta, tb);
137 /* a*0 = 0 or 0*b = 0:
138 calls computed_value recursive and returns the 0 with proper
140 if ((ta != tarval_bad) && (ta == get_mode_null(get_tarval_mode(ta))))
142 if ((tb != tarval_bad) && (tb == get_mode_null(get_tarval_mode(tb))))
149 * return the value of a floating point Quot
151 static tarval *computed_value_Quot(ir_node *n)
153 ir_node *a = get_Quot_left(n);
154 ir_node *b = get_Quot_right(n);
156 tarval *ta = value_of(a);
157 tarval *tb = value_of(b);
159 /* This was missing in original implementation. Why? */
160 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
161 if (tb != get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */
162 return tarval_quo(ta, tb);
168 * calculate the value of an integer Div of two nodes
169 * Special case: 0 / b
171 static tarval *do_computed_value_Div(ir_node *a, ir_node *b)
173 tarval *ta = value_of(a);
174 tarval *tb = value_of(b);
176 /* Compute c1 / c2 or 0 / a, a != 0 */
177 if (ta != tarval_bad) {
178 if ((tb != tarval_bad) && (tb != get_mode_null(get_irn_mode(b)))) /* div by zero: return tarval_bad */
179 return tarval_div(ta, tb);
180 else if (ta == get_mode_null(get_tarval_mode(ta))) /* 0 / b == 0 */
187 * return the value of an integer Div
189 static tarval *computed_value_Div(ir_node *n)
191 return do_computed_value_Div(get_Div_left(n), get_Div_right(n));
195 * calculate the value of an integer Mod of two nodes
196 * Special case: a % 1
198 static tarval *do_computed_value_Mod(ir_node *a, ir_node *b)
200 tarval *ta = value_of(a);
201 tarval *tb = value_of(b);
203 /* Compute c1 % c2 or a % 1 */
204 if (tb != tarval_bad) {
205 if ((ta != tarval_bad) && (tb != get_mode_null(get_tarval_mode(tb)))) /* div by zero: return tarval_bad */
206 return tarval_mod(ta, tb);
207 else if (tb == get_mode_one(get_tarval_mode(tb))) /* x mod 1 == 0 */
208 return get_mode_null(get_irn_mode(a));
215 * return the value of an integer Mod
217 static tarval *computed_value_Mod(ir_node *n)
219 return do_computed_value_Mod(get_Mod_left(n), get_Mod_right(n));
223 * return the value of an Abs
225 static tarval *computed_value_Abs(ir_node *n)
227 ir_node *a = get_Abs_op(n);
228 tarval *ta = value_of(a);
230 if (ta != tarval_bad)
231 return tarval_abs(ta);
237 * return the value of an And
238 * Special case: a & 0, 0 & b
240 static tarval *computed_value_And(ir_node *n)
242 ir_node *a = get_And_left(n);
243 ir_node *b = get_And_right(n);
245 tarval *ta = value_of(a);
246 tarval *tb = value_of(b);
248 if ((ta != tarval_bad) && (tb != tarval_bad)) {
249 return tarval_and (ta, tb);
253 if ( (classify_tarval ((v = ta)) == TV_CLASSIFY_NULL)
254 || (classify_tarval ((v = tb)) == TV_CLASSIFY_NULL)) {
262 * return the value of an Or
263 * Special case: a | 1...1, 1...1 | b
265 static tarval *computed_value_Or(ir_node *n)
267 ir_node *a = get_Or_left(n);
268 ir_node *b = get_Or_right(n);
270 tarval *ta = value_of(a);
271 tarval *tb = value_of(b);
273 if ((ta != tarval_bad) && (tb != tarval_bad)) {
274 return tarval_or (ta, tb);
277 if ( (classify_tarval ((v = ta)) == TV_CLASSIFY_ALL_ONE)
278 || (classify_tarval ((v = tb)) == TV_CLASSIFY_ALL_ONE)) {
286 * return the value of an Eor
288 static tarval *computed_value_Eor(ir_node *n)
290 ir_node *a = get_Eor_left(n);
291 ir_node *b = get_Eor_right(n);
296 return get_mode_null(get_irn_mode(n));
301 if ((ta != tarval_bad) && (tb != tarval_bad)) {
302 return tarval_eor (ta, tb);
308 * return the value of a Not
310 static tarval *computed_value_Not(ir_node *n)
312 ir_node *a = get_Not_op(n);
313 tarval *ta = value_of(a);
315 if (ta != tarval_bad)
316 return tarval_not(ta);
322 * return the value of a Shl
324 static tarval *computed_value_Shl(ir_node *n)
326 ir_node *a = get_Shl_left(n);
327 ir_node *b = get_Shl_right(n);
329 tarval *ta = value_of(a);
330 tarval *tb = value_of(b);
332 if ((ta != tarval_bad) && (tb != tarval_bad)) {
333 return tarval_shl (ta, tb);
339 * return the value of a Shr
341 static tarval *computed_value_Shr(ir_node *n)
343 ir_node *a = get_Shr_left(n);
344 ir_node *b = get_Shr_right(n);
346 tarval *ta = value_of(a);
347 tarval *tb = value_of(b);
349 if ((ta != tarval_bad) && (tb != tarval_bad)) {
350 return tarval_shr (ta, tb);
356 * return the value of a Shrs
358 static tarval *computed_value_Shrs(ir_node *n)
360 ir_node *a = get_Shrs_left(n);
361 ir_node *b = get_Shrs_right(n);
363 tarval *ta = value_of(a);
364 tarval *tb = value_of(b);
366 if ((ta != tarval_bad) && (tb != tarval_bad)) {
367 return tarval_shrs (ta, tb);
373 * return the value of a Rot
375 static tarval *computed_value_Rot(ir_node *n)
377 ir_node *a = get_Rot_left(n);
378 ir_node *b = get_Rot_right(n);
380 tarval *ta = value_of(a);
381 tarval *tb = value_of(b);
383 if ((ta != tarval_bad) && (tb != tarval_bad)) {
384 return tarval_rot (ta, tb);
390 * return the value of a Conv
392 static tarval *computed_value_Conv(ir_node *n)
394 ir_node *a = get_Conv_op(n);
395 tarval *ta = value_of(a);
397 if (ta != tarval_bad)
398 return tarval_convert_to(ta, get_irn_mode(n));
404 * return the value of a Proj(Cmp)
406 * This performs a first step of unreachable code elimination.
407 * Proj can not be computed, but folding a Cmp above the Proj here is
408 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
410 * There are several case where we can evaluate a Cmp node, see later.
412 static tarval *computed_value_Proj_Cmp(ir_node *n)
414 ir_node *a = get_Proj_pred(n);
415 ir_node *aa = get_Cmp_left(a);
416 ir_node *ab = get_Cmp_right(a);
417 long proj_nr = get_Proj_proj(n);
420 * BEWARE: a == a is NOT always True for floating Point values, as
421 * NaN != NaN is defined, so we must check this here.
424 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
427 /* This is a trick with the bits used for encoding the Cmp
428 Proj numbers, the following statement is not the same:
429 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
430 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
433 tarval *taa = value_of(aa);
434 tarval *tab = value_of(ab);
435 ir_mode *mode = get_irn_mode(aa);
438 * The predecessors of Cmp are target values. We can evaluate
441 if ((taa != tarval_bad) && (tab != tarval_bad)) {
442 /* strange checks... */
443 pn_Cmp flags = tarval_cmp(taa, tab);
444 if (flags != pn_Cmp_False) {
445 return new_tarval_from_long (proj_nr & flags, mode_b);
448 /* for integer values, we can check against MIN/MAX */
449 else if (mode_is_int(mode)) {
450 /* MIN <=/> x. This results in true/false. */
451 if (taa == get_mode_min(mode)) {
452 /* a compare with the MIN value */
453 if (proj_nr == pn_Cmp_Le)
454 return get_tarval_b_true();
455 else if (proj_nr == pn_Cmp_Gt)
456 return get_tarval_b_false();
458 /* x >=/< MIN. This results in true/false. */
460 if (tab == get_mode_min(mode)) {
461 /* a compare with the MIN value */
462 if (proj_nr == pn_Cmp_Ge)
463 return get_tarval_b_true();
464 else if (proj_nr == pn_Cmp_Lt)
465 return get_tarval_b_false();
467 /* MAX >=/< x. This results in true/false. */
468 else if (taa == get_mode_max(mode)) {
469 if (proj_nr == pn_Cmp_Ge)
470 return get_tarval_b_true();
471 else if (proj_nr == pn_Cmp_Lt)
472 return get_tarval_b_false();
474 /* x <=/> MAX. This results in true/false. */
475 else if (tab == get_mode_max(mode)) {
476 if (proj_nr == pn_Cmp_Le)
477 return get_tarval_b_true();
478 else if (proj_nr == pn_Cmp_Gt)
479 return get_tarval_b_false();
483 * The predecessors are Allocs or (void*)(0) constants. Allocs never
484 * return NULL, they raise an exception. Therefore we can predict
488 ir_node *aaa = skip_Id(skip_Proj(aa));
489 ir_node *aba = skip_Id(skip_Proj(ab));
491 if ( ( (/* aa is ProjP and aaa is Alloc */
492 (get_irn_op(aa) == op_Proj)
493 && (mode_is_reference(get_irn_mode(aa)))
494 && (get_irn_op(aaa) == op_Alloc))
495 && ( (/* ab is NULL */
496 (get_irn_op(ab) == op_Const)
497 && (mode_is_reference(get_irn_mode(ab)))
498 && (get_Const_tarval(ab) == get_mode_null(get_irn_mode(ab))))
499 || (/* ab is other Alloc */
500 (get_irn_op(ab) == op_Proj)
501 && (mode_is_reference(get_irn_mode(ab)))
502 && (get_irn_op(aba) == op_Alloc)
504 || (/* aa is NULL and aba is Alloc */
505 (get_irn_op(aa) == op_Const)
506 && (mode_is_reference(get_irn_mode(aa)))
507 && (get_Const_tarval(aa) == get_mode_null(get_irn_mode(aa)))
508 && (get_irn_op(ab) == op_Proj)
509 && (mode_is_reference(get_irn_mode(ab)))
510 && (get_irn_op(aba) == op_Alloc)))
512 return new_tarval_from_long(proj_nr & pn_Cmp_Ne, mode_b);
516 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
520 * return the value of a Proj, handle Proj(Cmp), Proj(Div), Proj(Mod), Proj(DivMod)
522 static tarval *computed_value_Proj(ir_node *n)
524 ir_node *a = get_Proj_pred(n);
527 switch (get_irn_opcode(a)) {
529 return computed_value_Proj_Cmp(n);
532 /* compute either the Div or the Mod part */
533 proj_nr = get_Proj_proj(n);
534 if (proj_nr == pn_DivMod_res_div)
535 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
536 else if (proj_nr == pn_DivMod_res_mod)
537 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
541 if (get_Proj_proj(n) == pn_Div_res)
542 return computed_value(a);
546 if (get_Proj_proj(n) == pn_Mod_res)
547 return computed_value(a);
557 * calculate the value of a Mux: can be evaluated, if the
558 * sel and the right input are known
560 static tarval *computed_value_Mux(ir_node *n)
562 ir_node *sel = get_Mux_sel(n);
563 tarval *ts = value_of(sel);
565 if (ts == get_tarval_b_true()) {
566 ir_node *v = get_Mux_true(n);
569 else if (ts == get_tarval_b_false()) {
570 ir_node *v = get_Mux_false(n);
577 * calculate the value of a Confirm: can be evaluated,
578 * if it has the form Confirm(x, '=', Const).
580 static tarval *computed_value_Confirm(ir_node *n)
582 return get_Confirm_cmp(n) == pn_Cmp_Eq ?
583 value_of(get_Confirm_bound(n)) : tarval_bad;
587 * If the parameter n can be computed, return its value, else tarval_bad.
588 * Performs constant folding.
590 * @param n The node this should be evaluated
592 tarval *computed_value(ir_node *n)
594 if (n->op->computed_value)
595 return n->op->computed_value(n);
600 * set the default computed_value evaluator
602 static ir_op *firm_set_default_computed_value(ir_op *op)
606 op->computed_value = computed_value_##a; \
633 op->computed_value = NULL;
641 /* returns 1 if the a and b are pointers to different locations. */
643 different_identity (ir_node *a, ir_node *b)
645 assert (mode_is_reference(get_irn_mode (a))
646 && mode_is_reference(get_irn_mode (b)));
648 if (get_irn_op (a) == op_Proj && get_irn_op(b) == op_Proj) {
649 ir_node *a1 = get_Proj_pred (a);
650 ir_node *b1 = get_Proj_pred (b);
651 if (a1 != b1 && get_irn_op (a1) == op_Alloc
652 && get_irn_op (b1) == op_Alloc)
660 * Returns a equivalent block for another block.
661 * If the block has only one predecessor, this is
662 * the equivalent one. If the only predecessor of a block is
663 * the block itself, this is a dead block.
665 * If both predecessors of a block are the branches of a binary
666 * Cond, the equivalent block is Cond's block.
668 * If all predecessors of a block are bad or lies in a dead
669 * block, the current block is dead as well.
671 * Note, that blocks are NEVER turned into Bad's, instead
672 * the dead_block flag is set. So, never test for is_Bad(block),
673 * always use is_dead_Block(block).
675 static ir_node *equivalent_node_Block(ir_node *n)
678 int n_preds = get_Block_n_cfgpreds(n);
680 /* The Block constructor does not call optimize, but mature_immBlock
681 calls the optimization. */
682 assert(get_Block_matured(n));
684 /* Straightening: a single entry Block following a single exit Block
685 can be merged, if it is not the Start block. */
686 /* !!! Beware, all Phi-nodes of n must have been optimized away.
687 This should be true, as the block is matured before optimize is called.
688 But what about Phi-cycles with the Phi0/Id that could not be resolved?
689 Remaining Phi nodes are just Ids. */
690 if ((n_preds == 1) && (get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp)) {
691 ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0));
692 if (predblock == oldn) {
693 /* Jmp jumps into the block it is in -- deal self cycle. */
694 n = set_Block_dead(n);
695 DBG_OPT_DEAD_BLOCK(oldn, n);
696 } else if (get_opt_control_flow_straightening()) {
698 DBG_OPT_STG(oldn, n);
701 else if ((n_preds == 1) &&
702 (get_irn_op(skip_Proj(get_Block_cfgpred(n, 0))) == op_Cond)) {
703 ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0));
704 if (predblock == oldn) {
705 /* Jmp jumps into the block it is in -- deal self cycle. */
706 n = set_Block_dead(n);
707 DBG_OPT_DEAD_BLOCK(oldn, n);
710 else if ((n_preds == 2) &&
711 (get_opt_control_flow_weak_simplification())) {
712 /* Test whether Cond jumps twice to this block
713 * The more general case which more than 2 predecessors is handles
714 * in optimize_cf(), we handle only this special case for speed here.
716 ir_node *a = get_Block_cfgpred(n, 0);
717 ir_node *b = get_Block_cfgpred(n, 1);
719 if ((get_irn_op(a) == op_Proj) &&
720 (get_irn_op(b) == op_Proj) &&
721 (get_Proj_pred(a) == get_Proj_pred(b)) &&
722 (get_irn_op(get_Proj_pred(a)) == op_Cond) &&
723 (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) {
724 /* Also a single entry Block following a single exit Block. Phis have
725 twice the same operand and will be optimized away. */
726 n = get_nodes_block(a);
727 DBG_OPT_IFSIM1(oldn, a, b, n);
730 else if (get_opt_unreachable_code() &&
731 (n != current_ir_graph->start_block) &&
732 (n != current_ir_graph->end_block) ) {
733 int i, n_cfg = get_Block_n_cfgpreds(n);
735 /* If all inputs are dead, this block is dead too, except if it is
736 the start or end block. This is a step of unreachable code
738 for (i = 0; i < n_cfg; i++) {
739 ir_node *pred = get_Block_cfgpred(n, i);
742 if (is_Bad(pred)) continue;
743 pred_blk = get_nodes_block(pred);
745 if (is_Block_dead(pred_blk)) continue;
748 /* really found a living input */
753 n = set_Block_dead(n);
760 * Returns a equivalent node for a Jmp, a Bad :-)
761 * Of course this only happens if the Block of the Jmp is Bad.
763 static ir_node *equivalent_node_Jmp(ir_node *n)
765 /* GL: Why not same for op_Raise?? */
766 /* unreachable code elimination */
767 if (is_Block_dead(get_nodes_block(n)))
773 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
774 See transform_node_Proj_Cond(). */
777 * optimize operations that are commutative and have neutral 0,
778 * so a op 0 = 0 op a = a.
780 static ir_node *equivalent_node_neutral_zero(ir_node *n)
784 ir_node *a = get_binop_left(n);
785 ir_node *b = get_binop_right(n);
790 /* After running compute_node there is only one constant predecessor.
791 Find this predecessors value and remember the other node: */
792 if ((tv = value_of(a)) != tarval_bad) {
794 } else if ((tv = value_of(b)) != tarval_bad) {
799 /* If this predecessors constant value is zero, the operation is
800 unnecessary. Remove it: */
801 if (classify_tarval (tv) == TV_CLASSIFY_NULL) {
804 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
810 #define equivalent_node_Eor equivalent_node_neutral_zero
813 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
815 * The second one looks strange, but this construct
816 * is used heavily in the LCC sources :-).
818 static ir_node *equivalent_node_Add(ir_node *n)
821 ir_node *left, *right;
823 n = equivalent_node_neutral_zero(n);
827 left = get_Add_left(n);
828 right = get_Add_right(n);
830 if (get_irn_op(left) == op_Sub) {
831 if (get_Sub_right(left) == right) {
834 n = get_Sub_left(left);
835 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
840 if (get_irn_op(right) == op_Sub) {
841 if (get_Sub_right(right) == left) {
844 n = get_Sub_left(left);
845 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
854 * optimize operations that are not commutative but have neutral 0 on left,
857 static ir_node *equivalent_node_left_zero(ir_node *n)
861 ir_node *a = get_binop_left(n);
862 ir_node *b = get_binop_right(n);
864 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
867 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
873 #define equivalent_node_Shl equivalent_node_left_zero
874 #define equivalent_node_Shr equivalent_node_left_zero
875 #define equivalent_node_Shrs equivalent_node_left_zero
876 #define equivalent_node_Rot equivalent_node_left_zero
879 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
881 * The second one looks strange, but this construct
882 * is used heavily in the LCC sources :-).
884 static ir_node *equivalent_node_Sub(ir_node *n)
888 ir_node *a = get_Sub_left(n);
889 ir_node *b = get_Sub_right(n);
891 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
894 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
896 else if (get_irn_op(a) == op_Add) {
897 ir_mode *mode = get_irn_mode(n);
899 if (mode_wrap_around(mode)) {
900 ir_node *left = get_Add_left(a);
901 ir_node *right = get_Add_right(a);
906 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
908 else if (right == b) {
911 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
921 * Optimize an "idempotent unary op", ie op(op(n)) = n.
923 * @fixme -(-a) == a, but might overflow two times.
924 * We handle it anyway here but the better way would be a
925 * flag. This would be needed for Pascal for instance.
927 static ir_node *equivalent_node_idempotent_unop(ir_node *n)
930 ir_node *pred = get_unop_op(n);
932 /* optimize symmetric unop */
933 if (get_irn_op(pred) == get_irn_op(n)) {
934 n = get_unop_op(pred);
935 DBG_OPT_ALGSIM2(oldn, pred, n);
940 /* Not(Not(x)) == x */
941 #define equivalent_node_Not equivalent_node_idempotent_unop
943 /* --x == x */ /* ??? Is this possible or can --x raise an
944 out of bounds exception if min =! max? */
945 #define equivalent_node_Minus equivalent_node_idempotent_unop
948 * Optimize a * 1 = 1 * a = a.
950 static ir_node *equivalent_node_Mul(ir_node *n)
954 ir_node *a = get_Mul_left(n);
955 ir_node *b = get_Mul_right(n);
957 /* Mul is commutative and has again an other neutral element. */
958 if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) {
960 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
961 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) {
963 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
969 * Optimize a / 1 = a.
971 static ir_node *equivalent_node_Div(ir_node *n)
973 ir_node *a = get_Div_left(n);
974 ir_node *b = get_Div_right(n);
976 /* Div is not commutative. */
977 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
978 /* Turn Div into a tuple (mem, bad, a) */
979 ir_node *mem = get_Div_mem(n);
980 turn_into_tuple(n, 3);
981 set_Tuple_pred(n, pn_Div_M, mem);
982 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
983 set_Tuple_pred(n, pn_Div_res, a);
989 * Optimize a / 1 = a.
991 static ir_node *equivalent_node_DivMod(ir_node *n)
993 ir_node *a = get_DivMod_left(n);
994 ir_node *b = get_DivMod_right(n);
996 /* Div is not commutative. */
997 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
998 /* Turn DivMod into a tuple (mem, bad, a, 0) */
999 ir_node *mem = get_Div_mem(n);
1000 ir_mode *mode = get_irn_mode(b);
1002 turn_into_tuple(n, 4);
1003 set_Tuple_pred(n, pn_DivMod_M, mem);
1004 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1005 set_Tuple_pred(n, pn_DivMod_res_div, a);
1006 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1012 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1014 static ir_node *equivalent_node_Or(ir_node *n)
1018 ir_node *a = get_Or_left(n);
1019 ir_node *b = get_Or_right(n);
1022 n = a; /* Or has it's own neutral element */
1023 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1024 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) {
1026 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1027 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
1029 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1036 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a.
1038 static ir_node *equivalent_node_And(ir_node *n)
1042 ir_node *a = get_And_left(n);
1043 ir_node *b = get_And_right(n);
1046 n = a; /* And has it's own neutral element */
1047 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1048 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) {
1050 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1051 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) {
1053 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1059 * Try to remove useless Conv's:
1061 static ir_node *equivalent_node_Conv(ir_node *n)
1064 ir_node *a = get_Conv_op(n);
1067 ir_mode *n_mode = get_irn_mode(n);
1068 ir_mode *a_mode = get_irn_mode(a);
1070 if (n_mode == a_mode) { /* No Conv necessary */
1072 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1073 } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */
1077 n_mode = get_irn_mode(n);
1078 b_mode = get_irn_mode(b);
1080 if (n_mode == b_mode) {
1081 if (n_mode == mode_b) {
1082 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1083 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1085 else if (mode_is_int(n_mode) || mode_is_character(n_mode)) {
1086 if (smaller_mode(b_mode, a_mode)){
1087 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1088 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1097 * A Cast may be removed if the type of the previous node
1098 * is already the type of the Cast.
1100 static ir_node *equivalent_node_Cast(ir_node *n) {
1102 ir_node *pred = get_Cast_op(n);
1104 if (get_irn_type(pred) == get_Cast_type(n)) {
1106 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1111 /* Several optimizations:
1112 - no Phi in start block.
1113 - remove Id operators that are inputs to Phi
1114 - fold Phi-nodes, iff they have only one predecessor except
1117 static ir_node *equivalent_node_Phi(ir_node *n)
1122 ir_node *block = NULL; /* to shutup gcc */
1123 ir_node *first_val = NULL; /* to shutup gcc */
1124 ir_node *scnd_val = NULL; /* to shutup gcc */
1126 if (!get_opt_normalize()) return n;
1128 n_preds = get_Phi_n_preds(n);
1130 block = get_nodes_block(n);
1131 /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
1132 assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
1133 if ((is_Block_dead(block)) || /* Control dead */
1134 (block == current_ir_graph->start_block)) /* There should be no Phi nodes */
1135 return new_Bad(); /* in the Start Block. */
1137 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1139 /* If the Block has a Bad pred, we also have one. */
1140 for (i = 0; i < n_preds; ++i)
1141 if (is_Bad(get_Block_cfgpred(block, i)))
1142 set_Phi_pred(n, i, new_Bad());
1144 /* Find first non-self-referencing input */
1145 for (i = 0; i < n_preds; ++i) {
1146 first_val = get_Phi_pred(n, i);
1147 if ( (first_val != n) /* not self pointer */
1149 && (! is_Bad(first_val))
1151 ) { /* value not dead */
1152 break; /* then found first value. */
1157 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1163 /* follow_Id () for rest of inputs, determine if any of these
1164 are non-self-referencing */
1165 while (++i < n_preds) {
1166 scnd_val = get_Phi_pred(n, i);
1167 if ( (scnd_val != n)
1168 && (scnd_val != first_val)
1170 && (! is_Bad(first_val))
1178 /* Fold, if no multiple distinct non-self-referencing inputs */
1180 DBG_OPT_PHI(oldn, n);
1182 /* skip the remaining Ids (done in get_Phi_pred). */
1183 /* superfluous, since we walk all to propagate Block's Bads.
1184 while (++i < n_preds) get_Phi_pred(n, i); */
1190 * optimize Proj(Tuple) and gigo() for ProjX in Bad block
1192 static ir_node *equivalent_node_Proj(ir_node *n)
1196 ir_node *a = get_Proj_pred(n);
1198 if ( get_irn_op(a) == op_Tuple) {
1199 /* Remove the Tuple/Proj combination. */
1200 if ( get_Proj_proj(n) <= get_Tuple_n_preds(a) ) {
1201 n = get_Tuple_pred(a, get_Proj_proj(n));
1202 DBG_OPT_TUPLE(oldn, a, n);
1204 assert(0); /* This should not happen! */
1207 } else if (get_irn_mode(n) == mode_X &&
1208 is_Block_dead(get_nodes_block(n))) {
1209 /* Remove dead control flow -- early gigo(). */
1218 static ir_node *equivalent_node_Id(ir_node *n)
1224 } while (get_irn_op(n) == op_Id);
1226 DBG_OPT_ID(oldn, n);
1233 static ir_node *equivalent_node_Mux(ir_node *n)
1235 ir_node *oldn = n, *sel = get_Mux_sel(n);
1236 tarval *ts = value_of(sel);
1238 /* Mux(true, f, t) == t */
1239 if (ts == tarval_b_true) {
1240 n = get_Mux_true(n);
1241 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1243 /* Mux(false, f, t) == f */
1244 else if (ts == tarval_b_false) {
1245 n = get_Mux_false(n);
1246 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1248 /* Mux(v, x, x) == x */
1249 else if (get_Mux_false(n) == get_Mux_true(n)) {
1250 n = get_Mux_true(n);
1251 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1253 else if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(get_irn_mode(n))) {
1254 ir_node *cmp = get_Proj_pred(sel);
1255 long proj_nr = get_Proj_proj(sel);
1256 ir_node *b = get_Mux_false(n);
1257 ir_node *a = get_Mux_true(n);
1260 * Note: normalization puts the constant on the right site,
1261 * so we check only one case.
1263 * Note further that these optimization work even for floating point
1264 * with NaN's because -NaN == NaN.
1265 * However, if +0 and -0 is handled differently, we cannot use the first one.
1267 if (get_irn_op(cmp) == op_Cmp && get_Cmp_left(cmp) == a) {
1268 if (classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
1269 /* Mux(a CMP 0, X, a) */
1270 if (get_irn_op(b) == op_Minus && get_Minus_op(b) == a) {
1271 /* Mux(a CMP 0, -a, a) */
1272 if (proj_nr == pn_Cmp_Eq) {
1273 /* Mux(a == 0, -a, a) ==> -a */
1275 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1277 else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1278 /* Mux(a != 0, -a, a) ==> a */
1280 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1283 else if (classify_Const(b) == CNST_NULL) {
1284 /* Mux(a CMP 0, 0, a) */
1285 if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1286 /* Mux(a != 0, 0, a) ==> a */
1288 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1290 else if (proj_nr == pn_Cmp_Eq) {
1291 /* Mux(a == 0, 0, a) ==> 0 */
1293 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1304 * Optimize -a CMP -b into b CMP a.
1305 * This works only for for modes where unary Minus
1307 * Note that two-complement integers can Overflow
1308 * so it will NOT work.
1310 static ir_node *equivalent_node_Cmp(ir_node *n)
1312 ir_node *left = get_Cmp_left(n);
1313 ir_node *right = get_Cmp_right(n);
1315 if (get_irn_op(left) == op_Minus && get_irn_op(right) == op_Minus &&
1316 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1317 left = get_Minus_op(left);
1318 right = get_Minus_op(right);
1319 set_Cmp_left(n, right);
1320 set_Cmp_right(n, left);
1326 * Remove Confirm nodes if setting is on.
1328 static ir_node *equivalent_node_Confirm(ir_node *n)
1330 if (get_Confirm_cmp(n) == pn_Cmp_Eq) {
1331 ir_node *bound = get_Confirm_bound(n);
1332 ir_op *op = get_irn_op(bound);
1335 * Optimize a rare case:
1336 * Confirm(x, '=', Const) ==> Const
1338 if (op == op_Const || op == op_SymConst)
1341 return get_opt_remove_Confirm() ? get_Confirm_value(n) : n;
1345 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1346 * perform no actual computation, as, e.g., the Id nodes. It does not create
1347 * new nodes. It is therefore safe to free n if the node returned is not n.
1348 * If a node returns a Tuple we can not just skip it. If the size of the
1349 * in array fits, we transform n into a tuple (e.g., Div).
1352 equivalent_node(ir_node *n)
1354 if (n->op->equivalent_node)
1355 return n->op->equivalent_node(n);
1360 * set the default equivalent node operation
1362 static ir_op *firm_set_default_equivalent_node(ir_op *op)
1366 op->equivalent_node = equivalent_node_##a; \
1395 op->equivalent_node = NULL;
1403 * Do node specific optimizations of nodes predecessors.
1406 optimize_preds(ir_node *n) {
1407 ir_node *a = NULL, *b = NULL;
1409 /* get the operands we will work on for simple cases. */
1411 a = get_binop_left(n);
1412 b = get_binop_right(n);
1413 } else if (is_unop(n)) {
1417 switch (get_irn_opcode(n)) {
1420 /* We don't want Cast as input to Cmp. */
1421 if (get_irn_op(a) == op_Cast) {
1425 if (get_irn_op(b) == op_Cast) {
1427 set_Cmp_right(n, b);
1436 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1437 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1438 * If possible, remove the Conv's.
1440 static ir_node *transform_node_AddSub(ir_node *n)
1442 ir_mode *mode = get_irn_mode(n);
1444 if (mode_is_reference(mode)) {
1445 ir_node *left = get_binop_left(n);
1446 ir_node *right = get_binop_right(n);
1447 int ref_bits = get_mode_size_bits(mode);
1449 if (get_irn_op(left) == op_Conv) {
1450 ir_mode *mode = get_irn_mode(left);
1451 int bits = get_mode_size_bits(mode);
1453 if (ref_bits == bits &&
1454 mode_is_int(mode) &&
1455 get_mode_arithmetic(mode) == irma_twos_complement) {
1456 ir_node *pre = get_Conv_op(left);
1457 ir_mode *pre_mode = get_irn_mode(pre);
1459 if (mode_is_int(pre_mode) &&
1460 get_mode_size_bits(pre_mode) == bits &&
1461 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1462 /* ok, this conv just changes to sign, moreover the calculation
1463 * is done with same number of bits as our address mode, so
1464 * we can ignore the conv as address calculation can be viewed
1465 * as either signed or unsigned
1467 set_binop_left(n, pre);
1472 if (get_irn_op(right) == op_Conv) {
1473 ir_mode *mode = get_irn_mode(right);
1474 int bits = get_mode_size_bits(mode);
1476 if (ref_bits == bits &&
1477 mode_is_int(mode) &&
1478 get_mode_arithmetic(mode) == irma_twos_complement) {
1479 ir_node *pre = get_Conv_op(right);
1480 ir_mode *pre_mode = get_irn_mode(pre);
1482 if (mode_is_int(pre_mode) &&
1483 get_mode_size_bits(pre_mode) == bits &&
1484 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1485 /* ok, this conv just changes to sign, moreover the calculation
1486 * is done with same number of bits as our address mode, so
1487 * we can ignore the conv as address calculation can be viewed
1488 * as either signed or unsigned
1490 set_binop_right(n, pre);
1499 * Do the AddSub optimization, then Transform Add(a,a) into Mul(a, 2)
1500 * if the mode is integer or float.
1501 * Transform Add(a,-b) into Sub(a,b).
1502 * Reassociation might fold this further.
1504 static ir_node *transform_node_Add(ir_node *n)
1509 n = transform_node_AddSub(n);
1511 mode = get_irn_mode(n);
1512 if (mode_is_num(mode)) {
1513 ir_node *a = get_Add_left(n);
1515 if (a == get_Add_right(n)) {
1516 ir_node *block = get_nodes_block(n);
1519 get_irn_dbg_info(n),
1523 new_r_Const_long(current_ir_graph, block, mode, 2),
1525 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
1528 ir_node *b = get_Add_right(n);
1530 if (get_irn_op(a) == op_Minus) {
1532 get_irn_dbg_info(n),
1538 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1540 else if (get_irn_op(b) == op_Minus) {
1542 get_irn_dbg_info(n),
1548 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1556 * Do the AddSub optimization, then Transform Sub(0,a) into Minus(a).
1558 static ir_node *transform_node_Sub(ir_node *n)
1563 n = transform_node_AddSub(n);
1565 mode = get_irn_mode(n);
1566 if (mode_is_num(mode) && (classify_Const(get_Sub_left(n)) == CNST_NULL)) {
1568 get_irn_dbg_info(n),
1573 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
1580 * Transform Mul(a,-1) into -a.
1581 * Do architecture dependent optimizations on Mul nodes
1583 static ir_node *transform_node_Mul(ir_node *n) {
1585 ir_mode *mode = get_irn_mode(n);
1587 if (mode_is_signed(mode)) {
1589 ir_node *a = get_Mul_left(n);
1590 ir_node *b = get_Mul_right(n);
1592 if (value_of(a) == get_mode_minus_one(mode))
1594 else if (value_of(b) == get_mode_minus_one(mode))
1597 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), r, mode);
1598 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
1602 return arch_dep_replace_mul_with_shifts(n);
1606 * transform a Div Node
1608 static ir_node *transform_node_Div(ir_node *n)
1610 tarval *tv = value_of(n);
1613 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
1615 if (tv != tarval_bad) {
1616 value = new_Const(get_tarval_mode(tv), tv);
1618 DBG_OPT_CSTEVAL(n, value);
1620 else /* Try architecture dependent optimization */
1621 value = arch_dep_replace_div_by_const(n);
1624 /* Turn Div into a tuple (mem, bad, value) */
1625 ir_node *mem = get_Div_mem(n);
1627 turn_into_tuple(n, 3);
1628 set_Tuple_pred(n, pn_Div_M, mem);
1629 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
1630 set_Tuple_pred(n, pn_Div_res, value);
1636 * transform a Mod node
1638 static ir_node *transform_node_Mod(ir_node *n)
1640 tarval *tv = value_of(n);
1643 /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */
1645 if (tv != tarval_bad) {
1646 value = new_Const(get_tarval_mode(tv), tv);
1648 DBG_OPT_CSTEVAL(n, value);
1650 else /* Try architecture dependent optimization */
1651 value = arch_dep_replace_mod_by_const(n);
1654 /* Turn Mod into a tuple (mem, bad, value) */
1655 ir_node *mem = get_Mod_mem(n);
1657 turn_into_tuple(n, 3);
1658 set_Tuple_pred(n, pn_Mod_M, mem);
1659 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
1660 set_Tuple_pred(n, pn_Mod_res, value);
1666 * transform a DivMod node
1668 static ir_node *transform_node_DivMod(ir_node *n)
1672 ir_node *a = get_DivMod_left(n);
1673 ir_node *b = get_DivMod_right(n);
1674 ir_mode *mode = get_irn_mode(a);
1675 tarval *ta = value_of(a);
1676 tarval *tb = value_of(b);
1678 if (!(mode_is_int(mode) && mode_is_int(get_irn_mode(b))))
1681 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
1683 if (tb != tarval_bad) {
1684 if (tb == get_mode_one(get_tarval_mode(tb))) {
1685 b = new_Const (mode, get_mode_null(mode));
1688 DBG_OPT_CSTEVAL(n, b);
1690 else if (ta != tarval_bad) {
1691 tarval *resa, *resb;
1692 resa = tarval_div (ta, tb);
1693 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
1694 Jmp for X result!? */
1695 resb = tarval_mod (ta, tb);
1696 if (resb == tarval_bad) return n; /* Causes exception! */
1697 a = new_Const (mode, resa);
1698 b = new_Const (mode, resb);
1701 DBG_OPT_CSTEVAL(n, a);
1702 DBG_OPT_CSTEVAL(n, b);
1704 else { /* Try architecture dependent optimization */
1705 arch_dep_replace_divmod_by_const(&a, &b, n);
1706 evaluated = a != NULL;
1708 } else if (ta == get_mode_null(mode)) {
1709 /* 0 / non-Const = 0 */
1714 if (evaluated) { /* replace by tuple */
1715 ir_node *mem = get_DivMod_mem(n);
1716 turn_into_tuple(n, 4);
1717 set_Tuple_pred(n, pn_DivMod_M, mem);
1718 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1719 set_Tuple_pred(n, pn_DivMod_res_div, a);
1720 set_Tuple_pred(n, pn_DivMod_res_mod, b);
1721 assert(get_nodes_block(n));
1728 * Optimize Abs(x) into x if x is Confirmed >= 0
1729 * Optimize Abs(x) into -x if x is Confirmed <= 0
1731 static ir_node *transform_node_Abs(ir_node *n)
1734 ir_node *a = get_Abs_op(n);
1735 value_classify sign = classify_value_sign(a);
1737 if (sign == VALUE_NEGATIVE) {
1738 ir_mode *mode = get_irn_mode(n);
1741 * We can replace the Abs by -x here.
1742 * We even could add a new Confirm here.
1744 * Note that -x would create a new node, so we could
1745 * not run it in the equivalent_node() context.
1747 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
1748 get_nodes_block(n), a, mode);
1750 DBG_OPT_CONFIRM(oldn, n);
1752 else if (sign == VALUE_POSITIVE) {
1753 /* n is positive, Abs is not needed */
1756 DBG_OPT_CONFIRM(oldn, n);
1763 * transform a Cond node
1765 static ir_node *transform_node_Cond(ir_node *n)
1767 /* Replace the Cond by a Jmp if it branches on a constant
1770 ir_node *a = get_Cond_selector(n);
1771 tarval *ta = value_of(a);
1773 if ((ta != tarval_bad) &&
1774 (get_irn_mode(a) == mode_b) &&
1775 (get_opt_unreachable_code())) {
1776 /* It's a boolean Cond, branching on a boolean constant.
1777 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
1778 jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n));
1779 turn_into_tuple(n, 2);
1780 if (ta == tarval_b_true) {
1781 set_Tuple_pred(n, pn_Cond_false, new_Bad());
1782 set_Tuple_pred(n, pn_Cond_true, jmp);
1784 set_Tuple_pred(n, pn_Cond_false, jmp);
1785 set_Tuple_pred(n, pn_Cond_true, new_Bad());
1787 /* We might generate an endless loop, so keep it alive. */
1788 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
1789 } else if ((ta != tarval_bad) &&
1790 (get_irn_mode(a) == mode_Iu) &&
1791 (get_Cond_kind(n) == dense) &&
1792 (get_opt_unreachable_code())) {
1793 /* I don't want to allow Tuples smaller than the biggest Proj.
1794 Also this tuple might get really big...
1795 I generate the Jmp here, and remember it in link. Link is used
1796 when optimizing Proj. */
1797 set_irn_link(n, new_r_Jmp(current_ir_graph, get_nodes_block(n)));
1798 /* We might generate an endless loop, so keep it alive. */
1799 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
1800 } else if ((get_irn_op(a) == op_Eor)
1801 && (get_irn_mode(a) == mode_b)
1802 && (classify_tarval(value_of(get_Eor_right(a))) == TV_CLASSIFY_ONE)) {
1803 /* The Eor is a negate. Generate a new Cond without the negate,
1804 simulate the negate by exchanging the results. */
1805 set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_block(n),
1807 } else if ((get_irn_op(a) == op_Not)
1808 && (get_irn_mode(a) == mode_b)) {
1809 /* A Not before the Cond. Generate a new Cond without the Not,
1810 simulate the Not by exchanging the results. */
1811 set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_block(n),
1820 static ir_node *transform_node_Eor(ir_node *n)
1823 ir_node *a = get_Eor_left(n);
1824 ir_node *b = get_Eor_right(n);
1825 ir_mode *mode = get_irn_mode(n);
1829 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n),
1830 mode, get_mode_null(mode));
1831 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
1833 else if ((mode == mode_b)
1834 && (get_irn_op(a) == op_Proj)
1835 && (get_irn_mode(a) == mode_b)
1836 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)
1837 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
1838 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
1839 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
1840 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
1842 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
1844 else if ((mode == mode_b)
1845 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) {
1846 /* The Eor is a Not. Replace it by a Not. */
1847 /* ????!!!Extend to bitfield 1111111. */
1848 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode_b);
1850 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
1857 * Transform a boolean Not.
1859 static ir_node *transform_node_Not(ir_node *n)
1862 ir_node *a = get_Not_op(n);
1864 if ( (get_irn_mode(n) == mode_b)
1865 && (get_irn_op(a) == op_Proj)
1866 && (get_irn_mode(a) == mode_b)
1867 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
1868 /* We negate a Cmp. The Cmp has the negated result anyways! */
1869 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
1870 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
1871 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
1878 * Transform a Cast_type(Const) into a new Const_type
1880 static ir_node *transform_node_Cast(ir_node *n) {
1882 ir_node *pred = get_Cast_op(n);
1883 type *tp = get_irn_type(n);
1885 if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) {
1886 n = new_rd_Const_type(NULL, current_ir_graph, get_nodes_block(pred), get_irn_mode(pred),
1887 get_Const_tarval(pred), tp);
1888 DBG_OPT_CSTEVAL(oldn, n);
1889 } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) {
1890 n = new_rd_SymConst_type(NULL, current_ir_graph, get_nodes_block(pred), get_SymConst_symbol(pred),
1891 get_SymConst_kind(pred), tp);
1892 DBG_OPT_CSTEVAL(oldn, n);
1899 * Transform a Proj(Div) with a non-zero value.
1900 * Removes the exceptions and routes the memory to the NoMem node.
1902 static ir_node *transform_node_Proj_Div(ir_node *proj)
1904 ir_node *n = get_Proj_pred(proj);
1905 ir_node *b = get_Div_right(n);
1908 if (value_not_zero(b)) {
1909 /* div(x, y) && y != 0 */
1910 proj_nr = get_Proj_proj(proj);
1912 /* this node may float */
1913 set_irn_pinned(n, op_pin_state_floats);
1915 if (proj_nr == pn_Div_X_except) {
1916 /* we found an exception handler, remove it */
1919 /* the memory Proj can be removed */
1920 ir_node *res = get_Div_mem(n);
1921 set_Div_mem(n, get_irg_no_mem(current_ir_graph));
1922 if (proj_nr == pn_Div_M)
1930 * Transform a Proj(Mod) with a non-zero value.
1931 * Removes the exceptions and routes the memory to the NoMem node.
1933 static ir_node *transform_node_Proj_Mod(ir_node *proj)
1935 ir_node *n = get_Proj_pred(proj);
1936 ir_node *b = get_Mod_right(n);
1939 if (value_not_zero(b)) {
1940 /* mod(x, y) && y != 0 */
1941 proj_nr = get_Proj_proj(proj);
1943 /* this node may float */
1944 set_irn_pinned(n, op_pin_state_floats);
1946 if (proj_nr == pn_Mod_X_except) {
1947 /* we found an exception handler, remove it */
1950 /* the memory Proj can be removed */
1951 ir_node *res = get_Mod_mem(n);
1952 set_Mod_mem(n, get_irg_no_mem(current_ir_graph));
1953 if (proj_nr == pn_Mod_M)
1961 * Transform a Proj(DivMod) with a non-zero value.
1962 * Removes the exceptions and routes the memory to the NoMem node.
1964 static ir_node *transform_node_Proj_DivMod(ir_node *proj)
1966 ir_node *n = get_Proj_pred(proj);
1967 ir_node *b = get_DivMod_right(n);
1970 if (value_not_zero(b)) {
1971 /* DivMod(x, y) && y != 0 */
1972 proj_nr = get_Proj_proj(proj);
1974 /* this node may float */
1975 set_irn_pinned(n, op_pin_state_floats);
1977 if (proj_nr == pn_DivMod_X_except) {
1978 /* we found an exception handler, remove it */
1982 /* the memory Proj can be removed */
1983 ir_node *res = get_DivMod_mem(n);
1984 set_DivMod_mem(n, get_irg_no_mem(current_ir_graph));
1985 if (proj_nr == pn_DivMod_M)
1993 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
1995 static ir_node *transform_node_Proj_Cond(ir_node *proj)
1997 if (get_opt_unreachable_code()) {
1998 ir_node *n = get_Proj_pred(proj);
1999 ir_node *b = get_Cond_selector(n);
2000 tarval *tb = value_of(b);
2002 if (tb != tarval_bad && mode_is_int(get_tarval_mode(tb))) {
2003 /* we have a constant switch */
2004 long num = get_Proj_proj(proj);
2006 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
2007 if (get_tarval_long(tb) == num) {
2008 /* Do NOT create a jump here, or we will have 2 control flow ops
2009 * in a block. This case is optimized away in optimize_cf(). */
2013 /* this case will NEVER be taken, kill it */
2023 * Normalizes and optimizes Cmp nodes.
2025 static ir_node *transform_node_Proj_Cmp(ir_node *proj)
2027 if (get_opt_reassociation()) {
2028 ir_node *n = get_Proj_pred(proj);
2029 ir_node *left = get_Cmp_left(n);
2030 ir_node *right = get_Cmp_right(n);
2034 ir_mode *mode = NULL;
2035 long proj_nr = get_Proj_proj(proj);
2038 * First step: normalize the compare op
2039 * by placing the constant on the right site
2040 * or moving the lower address node to the left.
2041 * We ignore the case that both are constants
2042 * this case should be optimized away.
2044 if (get_irn_op(right) == op_Const)
2046 else if (get_irn_op(left) == op_Const) {
2051 proj_nr = get_inversed_pnc(proj_nr);
2054 else if (left > right) {
2060 proj_nr = get_inversed_pnc(proj_nr);
2065 * Second step: Try to reduce the magnitude
2066 * of a constant. This may help to generate better code
2067 * later and may help to normalize more compares.
2068 * Of course this is only possible for integer values.
2071 mode = get_irn_mode(c);
2072 tv = get_Const_tarval(c);
2074 if (tv != tarval_bad) {
2075 /* the following optimization is possible on modes without Overflow
2076 * on Unary Minus or on == and !=:
2077 * -a CMP c ==> a swap(CMP) -c
2079 * Beware: for two-complement Overflow may occur, so only == and != can
2080 * be optimized, see this:
2081 * -MININT < 0 =/=> MININT > 0 !!!
2083 if (get_opt_constant_folding() && get_irn_op(left) == op_Minus &&
2084 (!mode_overflow_on_unary_Minus(mode) ||
2085 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
2086 left = get_Minus_op(left);
2087 tv = tarval_sub(get_mode_null(mode), tv);
2089 proj_nr = get_inversed_pnc(proj_nr);
2093 /* for integer modes, we have more */
2094 if (mode_is_int(mode)) {
2095 /* Ne includes Unordered which is not possible on integers.
2096 * However, frontends often use this wrong, so fix it here */
2097 if (proj_nr == pn_Cmp_Ne) {
2098 proj_nr = pn_Cmp_Lg;
2099 set_Proj_proj(proj, proj_nr);
2102 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
2103 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
2104 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
2105 tv = tarval_sub(tv, get_mode_one(mode));
2107 proj_nr ^= pn_Cmp_Eq;
2110 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
2111 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
2112 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
2113 tv = tarval_add(tv, get_mode_one(mode));
2115 proj_nr ^= pn_Cmp_Eq;
2119 /* the following reassociations work only for == and != */
2120 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
2122 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
2123 if (classify_tarval(tv) == TV_CLASSIFY_NULL && get_irn_op(left) == op_Sub) {
2124 right = get_Sub_right(left);
2125 left = get_Sub_left(left);
2127 tv = value_of(right);
2131 if (tv != tarval_bad) {
2132 ir_op *op = get_irn_op(left);
2134 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
2136 ir_node *c1 = get_Sub_right(left);
2137 tarval *tv2 = value_of(c1);
2139 if (tv2 != tarval_bad) {
2140 tv2 = tarval_add(tv, value_of(c1));
2142 if (tv2 != tarval_bad) {
2143 left = get_Sub_left(left);
2149 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
2150 else if (op == op_Add) {
2151 ir_node *a_l = get_Add_left(left);
2152 ir_node *a_r = get_Add_right(left);
2156 if (get_irn_op(a_l) == op_Const) {
2158 tv2 = value_of(a_l);
2162 tv2 = value_of(a_r);
2165 if (tv2 != tarval_bad) {
2166 tv2 = tarval_sub(tv, tv2);
2168 if (tv2 != tarval_bad) {
2175 /* -a == c ==> a == -c, -a != c ==> a != -c */
2176 else if (op == op_Minus) {
2177 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
2179 if (tv2 != tarval_bad) {
2180 left = get_Minus_op(left);
2187 /* the following reassociations work only for <= */
2188 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2189 if (tv != tarval_bad) {
2190 ir_op *op = get_irn_op(left);
2192 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
2202 ir_node *block = get_nodes_block(n);
2204 if (changed & 2) /* need a new Const */
2205 right = new_Const(mode, tv);
2207 /* create a new compare */
2208 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block,
2211 set_Proj_pred(proj, n);
2212 set_Proj_proj(proj, proj_nr);
2219 * Does all optimizations on nodes that must be done on it's Proj's
2220 * because of creating new nodes.
2222 static ir_node *transform_node_Proj(ir_node *proj)
2224 ir_node *n = get_Proj_pred(proj);
2226 switch (get_irn_opcode(n)) {
2228 return transform_node_Proj_Div(proj);
2231 return transform_node_Proj_Mod(proj);
2234 return transform_node_Proj_DivMod(proj);
2237 return transform_node_Proj_Cond(proj);
2240 return transform_node_Proj_Cmp(proj);
2243 /* should not happen, but if it does will be optimized away */
2244 return equivalent_node_Proj(proj);
2253 * returns the operands of a commutative bin-op, if one operand is
2254 * a const, it is returned as the second one.
2256 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c)
2258 ir_node *op_a = get_binop_left(binop);
2259 ir_node *op_b = get_binop_right(binop);
2261 assert(is_op_commutative(get_irn_op(binop)));
2263 if (get_irn_op(op_a) == op_Const) {
2274 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
2275 * Such pattern may arise in bitfield stores.
2277 * value c4 value c4 & c2
2278 * AND c3 AND c1 | c3
2283 static ir_node *transform_node_Or_bf_store(ir_node *or)
2287 ir_node *and_l, *c3;
2288 ir_node *value, *c4;
2289 ir_node *new_and, *new_const, *block;
2290 ir_mode *mode = get_irn_mode(or);
2292 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
2294 get_comm_Binop_Ops(or, &and, &c1);
2295 if ((get_irn_op(c1) != op_Const) || (get_irn_op(and) != op_And))
2298 get_comm_Binop_Ops(and, &or_l, &c2);
2299 if ((get_irn_op(c2) != op_Const) || (get_irn_op(or_l) != op_Or))
2302 get_comm_Binop_Ops(or_l, &and_l, &c3);
2303 if ((get_irn_op(c3) != op_Const) || (get_irn_op(and_l) != op_And))
2306 get_comm_Binop_Ops(and_l, &value, &c4);
2307 if (get_irn_op(c4) != op_Const)
2310 /* ok, found the pattern, check for conditions */
2311 assert(mode == get_irn_mode(and));
2312 assert(mode == get_irn_mode(or_l));
2313 assert(mode == get_irn_mode(and_l));
2315 tv1 = get_Const_tarval(c1);
2316 tv2 = get_Const_tarval(c2);
2317 tv3 = get_Const_tarval(c3);
2318 tv4 = get_Const_tarval(c4);
2320 tv = tarval_or(tv4, tv2);
2321 if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) {
2322 /* have at least one 0 at the same bit position */
2326 n_tv4 = tarval_not(tv4);
2327 if (tv3 != tarval_and(tv3, n_tv4)) {
2328 /* bit in the or_mask is outside the and_mask */
2332 n_tv2 = tarval_not(tv2);
2333 if (tv1 != tarval_and(tv1, n_tv2)) {
2334 /* bit in the or_mask is outside the and_mask */
2338 /* ok, all conditions met */
2339 block = get_nodes_block(or);
2341 new_and = new_r_And(current_ir_graph, block,
2342 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
2344 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
2346 set_Or_left(or, new_and);
2347 set_Or_right(or, new_const);
2349 /* check for more */
2350 return transform_node_Or_bf_store(or);
2354 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
2356 static ir_node *transform_node_Or_Rot(ir_node *or)
2358 ir_mode *mode = get_irn_mode(or);
2359 ir_node *shl, *shr, *block;
2360 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
2363 if (! mode_is_int(mode))
2366 shl = get_binop_left(or);
2367 shr = get_binop_right(or);
2369 if (get_irn_op(shl) == op_Shr) {
2370 if (get_irn_op(shr) != op_Shl)
2377 else if (get_irn_op(shl) != op_Shl)
2379 else if (get_irn_op(shr) != op_Shr)
2382 x = get_Shl_left(shl);
2383 if (x != get_Shr_left(shr))
2386 c1 = get_Shl_right(shl);
2387 c2 = get_Shr_right(shr);
2388 if (get_irn_op(c1) == op_Const && get_irn_op(c2) == op_Const) {
2389 tv1 = get_Const_tarval(c1);
2390 if (! tarval_is_long(tv1))
2393 tv2 = get_Const_tarval(c2);
2394 if (! tarval_is_long(tv2))
2397 if (get_tarval_long(tv1) + get_tarval_long(tv2)
2398 != get_mode_size_bits(mode))
2401 /* yet, condition met */
2402 block = get_nodes_block(or);
2404 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
2406 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
2409 else if (get_irn_op(c1) == op_Sub) {
2413 if (get_Sub_right(sub) != v)
2416 c1 = get_Sub_left(sub);
2417 if (get_irn_op(c1) != op_Const)
2420 tv1 = get_Const_tarval(c1);
2421 if (! tarval_is_long(tv1))
2424 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
2427 /* yet, condition met */
2428 block = get_nodes_block(or);
2430 /* a Rot right is not supported, so use a rot left */
2431 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
2433 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
2436 else if (get_irn_op(c2) == op_Sub) {
2440 c1 = get_Sub_left(sub);
2441 if (get_irn_op(c1) != op_Const)
2444 tv1 = get_Const_tarval(c1);
2445 if (! tarval_is_long(tv1))
2448 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
2451 /* yet, condition met */
2452 block = get_nodes_block(or);
2455 n = new_r_Rot(current_ir_graph, block, x, v, mode);
2457 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
2467 static ir_node *transform_node_Or(ir_node *or)
2469 or = transform_node_Or_bf_store(or);
2470 or = transform_node_Or_Rot(or);
2476 static ir_node *transform_node(ir_node *n);
2479 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl.
2481 * Should be moved to reassociation?
2483 static ir_node *transform_node_shift(ir_node *n)
2485 ir_node *left, *right;
2486 tarval *tv1, *tv2, *res;
2488 int modulo_shf, flag;
2490 left = get_binop_left(n);
2492 /* different operations */
2493 if (get_irn_op(left) != get_irn_op(n))
2496 right = get_binop_right(n);
2497 tv1 = value_of(right);
2498 if (tv1 == tarval_bad)
2501 tv2 = value_of(get_binop_right(left));
2502 if (tv2 == tarval_bad)
2505 res = tarval_add(tv1, tv2);
2507 /* beware: a simple replacement works only, if res < modulo shift */
2508 mode = get_irn_mode(n);
2512 modulo_shf = get_mode_modulo_shift(mode);
2513 if (modulo_shf > 0) {
2514 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
2516 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
2523 /* ok, we can replace it */
2524 ir_node *in[2], *irn, *block = get_nodes_block(n);
2526 in[0] = get_binop_left(left);
2527 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
2529 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
2531 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
2533 return transform_node(irn);
2538 #define transform_node_Shr transform_node_shift
2539 #define transform_node_Shrs transform_node_shift
2540 #define transform_node_Shl transform_node_shift
2543 * Remove dead blocks in keep alive list. We do not generate a new End node.
2545 static ir_node *transform_node_End(ir_node *n) {
2546 int i, n_keepalives = get_End_n_keepalives(n);
2548 for (i = 0; i < n_keepalives; ++i) {
2549 ir_node *ka = get_End_keepalive(n, i);
2550 if (is_Block(ka) && is_Block_dead(ka))
2551 set_End_keepalive(n, i, new_Bad());
2557 * Optimize a Mux into some simpler cases.
2559 static ir_node *transform_node_Mux(ir_node *n)
2561 ir_node *oldn = n, *sel = get_Mux_sel(n);
2562 ir_mode *mode = get_irn_mode(n);
2564 if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(mode)) {
2565 ir_node *cmp = get_Proj_pred(sel);
2566 long proj_nr = get_Proj_proj(sel);
2567 ir_node *f = get_Mux_false(n);
2568 ir_node *t = get_Mux_true(n);
2570 if (get_irn_op(cmp) == op_Cmp && classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
2571 ir_node *block = get_nodes_block(n);
2574 * Note: normalization puts the constant on the right site,
2575 * so we check only one case.
2577 * Note further that these optimization work even for floating point
2578 * with NaN's because -NaN == NaN.
2579 * However, if +0 and -0 is handled differently, we cannot use the first one.
2581 if (get_irn_op(f) == op_Minus &&
2582 get_Minus_op(f) == t &&
2583 get_Cmp_left(cmp) == t) {
2585 if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
2586 /* Mux(a >=/> 0, -a, a) ==> Abs(a) */
2587 n = new_rd_Abs(get_irn_dbg_info(n),
2591 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2594 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2595 /* Mux(a <=/< 0, -a, a) ==> Minus(Abs(a)) */
2596 n = new_rd_Abs(get_irn_dbg_info(n),
2600 n = new_rd_Minus(get_irn_dbg_info(n),
2605 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2609 else if (get_irn_op(t) == op_Minus &&
2610 get_Minus_op(t) == f &&
2611 get_Cmp_left(cmp) == f) {
2613 if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2614 /* Mux(a <=/< 0, a, -a) ==> Abs(a) */
2615 n = new_rd_Abs(get_irn_dbg_info(n),
2619 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2622 else if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
2623 /* Mux(a >=/> 0, a, -a) ==> Minus(Abs(a)) */
2624 n = new_rd_Abs(get_irn_dbg_info(n),
2628 n = new_rd_Minus(get_irn_dbg_info(n),
2633 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2638 if (mode_is_int(mode) && mode_is_signed(mode) &&
2639 get_mode_arithmetic(mode) == irma_twos_complement) {
2640 ir_node *x = get_Cmp_left(cmp);
2642 /* the following optimization works only with signed integer two-complement mode */
2644 if (mode == get_irn_mode(x)) {
2646 * FIXME: this restriction is two rigid, as it would still
2647 * work if mode(x) = Hs and mode == Is, but at least it removes
2650 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Le) &&
2651 classify_Const(t) == CNST_ALL_ONE &&
2652 classify_Const(f) == CNST_NULL) {
2654 * Mux(x:T </<= 0, 0, -1) -> Shrs(x, sizeof_bits(T) - 1)
2658 n = new_rd_Shrs(get_irn_dbg_info(n),
2659 current_ir_graph, block, x,
2660 new_r_Const_long(current_ir_graph, block, mode_Iu,
2661 get_mode_size_bits(mode) - 1),
2663 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
2666 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) &&
2667 classify_Const(t) == CNST_ONE &&
2668 classify_Const(f) == CNST_NULL) {
2670 * Mux(x:T >/>= 0, 0, 1) -> Shr(-x, sizeof_bits(T) - 1)
2674 n = new_rd_Shr(get_irn_dbg_info(n),
2675 current_ir_graph, block,
2676 new_r_Minus(current_ir_graph, block, x, mode),
2677 new_r_Const_long(current_ir_graph, block, mode_Iu,
2678 get_mode_size_bits(mode) - 1),
2680 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
2687 return arch_transform_node_Mux(n);
2691 * Tries several [inplace] [optimizing] transformations and returns an
2692 * equivalent node. The difference to equivalent_node() is that these
2693 * transformations _do_ generate new nodes, and thus the old node must
2694 * not be freed even if the equivalent node isn't the old one.
2696 static ir_node *transform_node(ir_node *n)
2698 if (n->op->transform_node)
2699 n = n->op->transform_node(n);
2704 * set the default transform node operation
2706 static ir_op *firm_set_default_transform_node(ir_op *op)
2710 op->transform_node = transform_node_##a; \
2734 op->transform_node = NULL;
2742 /* **************** Common Subexpression Elimination **************** */
2744 /** The size of the hash table used, should estimate the number of nodes
2746 #define N_IR_NODES 512
2748 /** Compares the attributes of two Const nodes. */
2749 static int node_cmp_attr_Const(ir_node *a, ir_node *b)
2751 return (get_Const_tarval(a) != get_Const_tarval(b))
2752 || (get_Const_type(a) != get_Const_type(b));
2755 /** Compares the attributes of two Proj nodes. */
2756 static int node_cmp_attr_Proj(ir_node *a, ir_node *b)
2758 return get_irn_proj_attr (a) != get_irn_proj_attr (b);
2761 /** Compares the attributes of two Filter nodes. */
2762 static int node_cmp_attr_Filter(ir_node *a, ir_node *b)
2764 return get_Filter_proj(a) != get_Filter_proj(b);
2767 /** Compares the attributes of two Alloc nodes. */
2768 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b)
2770 return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where)
2771 || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type);
2774 /** Compares the attributes of two Free nodes. */
2775 static int node_cmp_attr_Free(ir_node *a, ir_node *b)
2777 return (get_irn_free_attr(a).where != get_irn_free_attr(b).where)
2778 || (get_irn_free_attr(a).type != get_irn_free_attr(b).type);
2781 /** Compares the attributes of two SymConst nodes. */
2782 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b)
2784 return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num)
2785 || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p)
2786 || (get_irn_symconst_attr(a).tp != get_irn_symconst_attr(b).tp);
2789 /** Compares the attributes of two Call nodes. */
2790 static int node_cmp_attr_Call(ir_node *a, ir_node *b)
2792 return (get_irn_call_attr(a) != get_irn_call_attr(b));
2795 /** Compares the attributes of two Sel nodes. */
2796 static int node_cmp_attr_Sel(ir_node *a, ir_node *b)
2798 return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind)
2799 || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name)
2800 || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner)
2801 || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name)
2802 || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type);
2805 /** Compares the attributes of two Phi nodes. */
2806 static int node_cmp_attr_Phi(ir_node *a, ir_node *b)
2808 return get_irn_phi_attr (a) != get_irn_phi_attr (b);
2811 /** Compares the attributes of two Cast nodes. */
2812 static int node_cmp_attr_Cast(ir_node *a, ir_node *b)
2814 return get_Cast_type(a) != get_Cast_type(b);
2817 /** Compares the attributes of two Load nodes. */
2818 static int node_cmp_attr_Load(ir_node *a, ir_node *b)
2820 if (get_Load_volatility(a) == volatility_is_volatile ||
2821 get_Load_volatility(b) == volatility_is_volatile)
2822 /* NEVER do CSE on volatile Loads */
2825 return get_Load_mode(a) != get_Load_mode(b);
2828 /** Compares the attributes of two Store nodes. */
2829 static int node_cmp_attr_Store(ir_node *a, ir_node *b)
2831 /* NEVER do CSE on volatile Stores */
2832 return (get_Store_volatility(a) == volatility_is_volatile ||
2833 get_Store_volatility(b) == volatility_is_volatile);
2836 /** Compares the attributes of two Confirm nodes. */
2837 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b)
2839 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
2843 * set the default node attribute compare operation
2845 static ir_op *firm_set_default_node_cmp_attr(ir_op *op)
2849 op->node_cmp_attr = node_cmp_attr_##a; \
2867 op->node_cmp_attr = NULL;
2875 * Compare function for two nodes in the hash table. Gets two
2876 * nodes as parameters. Returns 0 if the nodes are a cse.
2879 vt_cmp (const void *elt, const void *key)
2887 if (a == b) return 0;
2889 if ((get_irn_op(a) != get_irn_op(b)) ||
2890 (get_irn_mode(a) != get_irn_mode(b))) return 1;
2892 /* compare if a's in and b's in are of equal length */
2893 irn_arity_a = get_irn_intra_arity (a);
2894 if (irn_arity_a != get_irn_intra_arity(b))
2897 /* for block-local cse and op_pin_state_pinned nodes: */
2898 if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) {
2899 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
2903 /* compare a->in[0..ins] with b->in[0..ins] */
2904 for (i = 0; i < irn_arity_a; i++)
2905 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
2909 * here, we already now that the nodes are identical except their
2912 if (a->op->node_cmp_attr)
2913 return a->op->node_cmp_attr(a, b);
2919 * Calculate a hash value of a node.
2922 ir_node_hash (ir_node *node)
2927 if (node->op == op_Const) {
2928 /* special value for const, as they only differ in their tarval. */
2929 h = HASH_PTR(node->attr.con.tv);
2930 h = 9*h + HASH_PTR(get_irn_mode(node));
2931 } else if (node->op == op_SymConst) {
2932 /* special value for const, as they only differ in their symbol. */
2933 h = HASH_PTR(node->attr.i.sym.type_p);
2934 h = 9*h + HASH_PTR(get_irn_mode(node));
2937 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
2938 h = irn_arity = get_irn_intra_arity(node);
2940 /* consider all in nodes... except the block if not a control flow. */
2941 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
2942 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
2946 h = 9*h + HASH_PTR(get_irn_mode(node));
2948 h = 9*h + HASH_PTR(get_irn_op(node));
2955 new_identities(void) {
2956 return new_pset(vt_cmp, N_IR_NODES);
2960 del_identities(pset *value_table) {
2961 del_pset(value_table);
2965 * Return the canonical node computing the same value as n.
2966 * Looks up the node in a hash table.
2968 * For Const nodes this is performed in the constructor, too. Const
2969 * nodes are extremely time critical because of their frequent use in
2970 * constant string arrays.
2972 static INLINE ir_node *
2973 identify (pset *value_table, ir_node *n)
2977 if (!value_table) return n;
2979 if (get_opt_reassociation()) {
2980 if (is_op_commutative(get_irn_op(n))) {
2981 ir_node *l = get_binop_left(n);
2982 ir_node *r = get_binop_right(n);
2984 /* for commutative operators perform a OP b == b OP a */
2986 set_binop_left(n, r);
2987 set_binop_right(n, l);
2992 o = pset_find (value_table, n, ir_node_hash (n));
3001 * During construction we set the op_pin_state_pinned flag in the graph right when the
3002 * optimization is performed. The flag turning on procedure global cse could
3003 * be changed between two allocations. This way we are safe.
3005 static INLINE ir_node *
3006 identify_cons (pset *value_table, ir_node *n) {
3009 n = identify(value_table, n);
3010 if (get_irn_n(old, -1) != get_irn_n(n, -1))
3011 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3016 * Return the canonical node computing the same value as n.
3017 * Looks up the node in a hash table, enters it in the table
3018 * if it isn't there yet.
3021 identify_remember (pset *value_table, ir_node *n)
3025 if (!value_table) return n;
3027 if (get_opt_reassociation()) {
3028 if (is_op_commutative(get_irn_op(n))) {
3029 ir_node *l = get_binop_left(n);
3030 ir_node *r = get_binop_right(n);
3032 /* for commutative operators perform a OP b == b OP a */
3034 set_binop_left(n, r);
3035 set_binop_right(n, l);
3040 /* lookup or insert in hash table with given hash key. */
3041 o = pset_insert (value_table, n, ir_node_hash (n));
3051 add_identities (pset *value_table, ir_node *node) {
3052 if (get_opt_cse() && (get_irn_opcode(node) != iro_Block))
3053 identify_remember (value_table, node);
3057 * garbage in, garbage out. If a node has a dead input, i.e., the
3058 * Bad node is input to the node, return the Bad node.
3060 static INLINE ir_node *
3061 gigo (ir_node *node)
3064 ir_op* op = get_irn_op(node);
3066 /* remove garbage blocks by looking at control flow that leaves the block
3067 and replacing the control flow by Bad. */
3068 if (get_irn_mode(node) == mode_X) {
3069 ir_node *block = get_nodes_block(node);
3070 if (!get_Block_matured(block)) return node; /* Don't optimize nodes in immature blocks. */
3071 if (op == op_End) return node; /* Don't optimize End, may have Bads. */
3073 if (get_irn_op(block) == op_Block && get_Block_matured(block)) {
3074 irn_arity = get_irn_arity(block);
3075 for (i = 0; i < irn_arity; i++) {
3076 if (!is_Bad(get_irn_n(block, i))) break;
3078 if (i == irn_arity) return new_Bad();
3082 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
3083 blocks predecessors is dead. */
3084 if ( op != op_Block && op != op_Phi && op != op_Tuple) {
3085 irn_arity = get_irn_arity(node);
3087 if (is_Block_dead(get_nodes_block(node)))
3090 for (i = 0; i < irn_arity; i++) {
3091 if (is_Bad(get_irn_n(node, i))) {
3097 /* With this code we violate the agreement that local_optimize
3098 only leaves Bads in Block, Phi and Tuple nodes. */
3099 /* If Block has only Bads as predecessors it's garbage. */
3100 /* If Phi has only Bads as predecessors it's garbage. */
3101 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
3102 irn_arity = get_irn_arity(node);
3103 for (i = 0; i < irn_arity; i++) {
3104 if (!is_Bad(get_irn_n(node, i))) break;
3106 if (i == irn_arity) node = new_Bad();
3114 * These optimizations deallocate nodes from the obstack.
3115 * It can only be called if it is guaranteed that no other nodes
3116 * reference this one, i.e., right after construction of a node.
3119 optimize_node(ir_node *n)
3123 opcode iro = get_irn_opcode(n);
3125 /* Always optimize Phi nodes: part of the construction. */
3126 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
3128 /* constant expression evaluation / constant folding */
3129 if (get_opt_constant_folding()) {
3130 /* neither constants nor Tuple values can be evaluated */
3131 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
3132 /* try to evaluate */
3133 tv = computed_value(n);
3134 if (tv != tarval_bad) {
3136 type *old_tp = get_irn_type(n);
3137 int i, arity = get_irn_arity(n);
3141 * Try to recover the type of the new expression.
3143 for (i = 0; i < arity && !old_tp; ++i)
3144 old_tp = get_irn_type(get_irn_n(n, i));
3147 * we MUST copy the node here temporary, because it's still needed
3148 * for DBG_OPT_CSTEVAL
3150 node_size = offsetof(ir_node, attr) + n->op->attr_size;
3151 oldn = alloca(node_size);
3153 memcpy(oldn, n, node_size);
3154 CLONE_ARR_A(ir_node *, oldn->in, n->in);
3156 /* ARG, copy the in array, we need it for statistics */
3157 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
3159 /* note the inplace edges module */
3160 edges_node_deleted(n, current_ir_graph);
3162 /* evaluation was successful -- replace the node. */
3163 obstack_free(current_ir_graph->obst, n);
3164 nw = new_Const(get_tarval_mode (tv), tv);
3166 if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv))
3167 set_Const_type(nw, old_tp);
3168 DBG_OPT_CSTEVAL(oldn, nw);
3174 /* remove unnecessary nodes */
3175 if (get_opt_constant_folding() ||
3176 (iro == iro_Phi) || /* always optimize these nodes. */
3178 (iro == iro_Proj) ||
3179 (iro == iro_Block) ) /* Flags tested local. */
3180 n = equivalent_node (n);
3182 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
3184 /* Common Subexpression Elimination.
3186 * Checks whether n is already available.
3187 * The block input is used to distinguish different subexpressions. Right
3188 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
3189 * subexpressions within a block.
3192 n = identify_cons (current_ir_graph->value_table, n);
3195 edges_node_deleted(oldn, current_ir_graph);
3197 /* We found an existing, better node, so we can deallocate the old node. */
3198 obstack_free (current_ir_graph->obst, oldn);
3203 /* Some more constant expression evaluation that does not allow to
3205 iro = get_irn_opcode(n);
3206 if (get_opt_constant_folding() ||
3207 (iro == iro_Cond) ||
3208 (iro == iro_Proj) ||
3209 (iro == iro_Sel)) /* Flags tested local. */
3210 n = transform_node (n);
3212 /* Remove nodes with dead (Bad) input.
3213 Run always for transformation induced Bads. */
3216 /* Now we have a legal, useful node. Enter it in hash table for cse */
3217 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
3218 n = identify_remember (current_ir_graph->value_table, n);
3226 * These optimizations never deallocate nodes (in place). This can cause dead
3227 * nodes lying on the obstack. Remove these by a dead node elimination,
3228 * i.e., a copying garbage collection.
3231 optimize_in_place_2 (ir_node *n)
3235 opcode iro = get_irn_opcode(n);
3237 if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n;
3239 /* constant expression evaluation / constant folding */
3240 if (get_opt_constant_folding()) {
3241 /* neither constants nor Tuple values can be evaluated */
3242 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
3243 /* try to evaluate */
3244 tv = computed_value(n);
3245 if (tv != tarval_bad) {
3246 /* evaluation was successful -- replace the node. */
3247 type *old_tp = get_irn_type(n);
3248 int i, arity = get_irn_arity(n);
3251 * Try to recover the type of the new expression.
3253 for (i = 0; i < arity && !old_tp; ++i)
3254 old_tp = get_irn_type(get_irn_n(n, i));
3256 n = new_Const(get_tarval_mode(tv), tv);
3258 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
3259 set_Const_type(n, old_tp);
3261 DBG_OPT_CSTEVAL(oldn, n);
3267 /* remove unnecessary nodes */
3268 if (get_opt_constant_folding() ||
3269 (iro == iro_Phi) || /* always optimize these nodes. */
3270 (iro == iro_Id) || /* ... */
3271 (iro == iro_Proj) || /* ... */
3272 (iro == iro_Block) ) /* Flags tested local. */
3273 n = equivalent_node(n);
3275 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
3277 /** common subexpression elimination **/
3278 /* Checks whether n is already available. */
3279 /* The block input is used to distinguish different subexpressions. Right
3280 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
3281 subexpressions within a block. */
3282 if (get_opt_cse()) {
3283 n = identify(current_ir_graph->value_table, n);
3286 /* Some more constant expression evaluation. */
3287 iro = get_irn_opcode(n);
3288 if (get_opt_constant_folding() ||
3289 (iro == iro_Cond) ||
3290 (iro == iro_Proj) ||
3291 (iro == iro_Sel)) /* Flags tested local. */
3292 n = transform_node(n);
3294 /* Remove nodes with dead (Bad) input.
3295 Run always for transformation induced Bads. */
3298 /* Now we can verify the node, as it has no dead inputs any more. */
3301 /* Now we have a legal, useful node. Enter it in hash table for cse.
3302 Blocks should be unique anyways. (Except the successor of start:
3303 is cse with the start block!) */
3304 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
3305 n = identify_remember(current_ir_graph->value_table, n);
3311 * Wrapper for external use, set proper status bits after optimization.
3314 optimize_in_place (ir_node *n)
3316 /* Handle graph state */
3317 assert(get_irg_phase_state(current_ir_graph) != phase_building);
3319 if (get_opt_global_cse())
3320 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3321 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
3322 set_irg_outs_inconsistent(current_ir_graph);
3324 /* Maybe we could also test whether optimizing the node can
3325 change the control graph. */
3326 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
3327 set_irg_dom_inconsistent(current_ir_graph);
3328 return optimize_in_place_2 (n);
3332 * set the default ir op operations
3334 ir_op *firm_set_default_operations(ir_op *op)
3336 op = firm_set_default_computed_value(op);
3337 op = firm_set_default_equivalent_node(op);
3338 op = firm_set_default_transform_node(op);
3339 op = firm_set_default_node_cmp_attr(op);
3340 op = firm_set_default_get_type(op);