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_Block_cfgpred_block(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(skip_Proj(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.
802 * Beware: If n is a Add, the mode of on and n might be different
803 * which happens in this rare construction: NULL + 3.
804 * Then, a Conv would be needed which we cannot include here.
806 if (classify_tarval (tv) == TV_CLASSIFY_NULL) {
807 if (get_irn_mode(on) == get_irn_mode(n)) {
810 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
817 #define equivalent_node_Eor equivalent_node_neutral_zero
820 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
822 * The second one looks strange, but this construct
823 * is used heavily in the LCC sources :-).
825 * Beware: The Mode of an Add may be different than the mode of its
826 * predecessors, so we could not return a predecessors in all cases.
828 static ir_node *equivalent_node_Add(ir_node *n)
831 ir_node *left, *right;
833 n = equivalent_node_neutral_zero(n);
837 left = get_Add_left(n);
838 right = get_Add_right(n);
840 if (get_irn_op(left) == op_Sub) {
841 if (get_Sub_right(left) == right) {
844 n = get_Sub_left(left);
845 if (get_irn_mode(oldn) == get_irn_mode(n)) {
846 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
851 if (get_irn_op(right) == op_Sub) {
852 if (get_Sub_right(right) == left) {
855 n = get_Sub_left(right);
856 if (get_irn_mode(oldn) == get_irn_mode(n)) {
857 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
866 * optimize operations that are not commutative but have neutral 0 on left,
869 static ir_node *equivalent_node_left_zero(ir_node *n)
873 ir_node *a = get_binop_left(n);
874 ir_node *b = get_binop_right(n);
876 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
879 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
885 #define equivalent_node_Shl equivalent_node_left_zero
886 #define equivalent_node_Shr equivalent_node_left_zero
887 #define equivalent_node_Shrs equivalent_node_left_zero
888 #define equivalent_node_Rot equivalent_node_left_zero
891 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
893 * The second one looks strange, but this construct
894 * is used heavily in the LCC sources :-).
896 * Beware: The Mode of a Sub may be different than the mode of its
897 * predecessors, so we could not return a predecessors in all cases.
899 static ir_node *equivalent_node_Sub(ir_node *n)
903 ir_node *a = get_Sub_left(n);
904 ir_node *b = get_Sub_right(n);
906 /* Beware: modes might be different */
907 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
908 if (get_irn_mode(n) == get_irn_mode(a)) {
911 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
914 else if (get_irn_op(a) == op_Add) {
915 ir_mode *mode = get_irn_mode(n);
917 if (mode_wrap_around(mode)) {
918 ir_node *left = get_Add_left(a);
919 ir_node *right = get_Add_right(a);
922 if (get_irn_mode(n) == get_irn_mode(right)) {
924 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
927 else if (right == b) {
928 if (get_irn_mode(n) == get_irn_mode(left)) {
930 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
941 * Optimize an "idempotent unary op", ie op(op(n)) = n.
943 * @fixme -(-a) == a, but might overflow two times.
944 * We handle it anyway here but the better way would be a
945 * flag. This would be needed for Pascal for instance.
947 static ir_node *equivalent_node_idempotent_unop(ir_node *n)
950 ir_node *pred = get_unop_op(n);
952 /* optimize symmetric unop */
953 if (get_irn_op(pred) == get_irn_op(n)) {
954 n = get_unop_op(pred);
955 DBG_OPT_ALGSIM2(oldn, pred, n);
960 /* Not(Not(x)) == x */
961 #define equivalent_node_Not equivalent_node_idempotent_unop
963 /* --x == x */ /* ??? Is this possible or can --x raise an
964 out of bounds exception if min =! max? */
965 #define equivalent_node_Minus equivalent_node_idempotent_unop
968 * Optimize a * 1 = 1 * a = a.
970 static ir_node *equivalent_node_Mul(ir_node *n)
974 ir_node *a = get_Mul_left(n);
975 ir_node *b = get_Mul_right(n);
977 /* Mul is commutative and has again an other neutral element. */
978 if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) {
980 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
981 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) {
983 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
989 * Optimize a / 1 = a.
991 static ir_node *equivalent_node_Div(ir_node *n)
993 ir_node *a = get_Div_left(n);
994 ir_node *b = get_Div_right(n);
996 /* Div is not commutative. */
997 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
998 /* Turn Div into a tuple (mem, bad, a) */
999 ir_node *mem = get_Div_mem(n);
1000 turn_into_tuple(n, 3);
1001 set_Tuple_pred(n, pn_Div_M, mem);
1002 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
1003 set_Tuple_pred(n, pn_Div_res, a);
1009 * Optimize a / 1 = a.
1011 static ir_node *equivalent_node_DivMod(ir_node *n)
1013 ir_node *a = get_DivMod_left(n);
1014 ir_node *b = get_DivMod_right(n);
1016 /* Div is not commutative. */
1017 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1018 /* Turn DivMod into a tuple (mem, bad, a, 0) */
1019 ir_node *mem = get_Div_mem(n);
1020 ir_mode *mode = get_irn_mode(b);
1022 turn_into_tuple(n, 4);
1023 set_Tuple_pred(n, pn_DivMod_M, mem);
1024 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1025 set_Tuple_pred(n, pn_DivMod_res_div, a);
1026 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1032 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1034 static ir_node *equivalent_node_Or(ir_node *n)
1038 ir_node *a = get_Or_left(n);
1039 ir_node *b = get_Or_right(n);
1042 n = a; /* Or has it's own neutral element */
1043 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1044 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) {
1046 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1047 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
1049 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1056 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a.
1058 static ir_node *equivalent_node_And(ir_node *n)
1062 ir_node *a = get_And_left(n);
1063 ir_node *b = get_And_right(n);
1066 n = a; /* And has it's own neutral element */
1067 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1068 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) {
1070 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1071 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) {
1073 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1079 * Try to remove useless Conv's:
1081 static ir_node *equivalent_node_Conv(ir_node *n)
1084 ir_node *a = get_Conv_op(n);
1087 ir_mode *n_mode = get_irn_mode(n);
1088 ir_mode *a_mode = get_irn_mode(a);
1090 if (n_mode == a_mode) { /* No Conv necessary */
1092 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1093 } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */
1097 n_mode = get_irn_mode(n);
1098 b_mode = get_irn_mode(b);
1100 if (n_mode == b_mode) {
1101 if (n_mode == mode_b) {
1102 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1103 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1105 else if (mode_is_int(n_mode) || mode_is_character(n_mode)) {
1106 if (smaller_mode(b_mode, a_mode)){
1107 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1108 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1117 * A Cast may be removed if the type of the previous node
1118 * is already the type of the Cast.
1120 static ir_node *equivalent_node_Cast(ir_node *n) {
1122 ir_node *pred = get_Cast_op(n);
1124 if (get_irn_type(pred) == get_Cast_type(n)) {
1126 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1131 /* Several optimizations:
1132 - no Phi in start block.
1133 - remove Id operators that are inputs to Phi
1134 - fold Phi-nodes, iff they have only one predecessor except
1137 static ir_node *equivalent_node_Phi(ir_node *n)
1142 ir_node *block = NULL; /* to shutup gcc */
1143 ir_node *first_val = NULL; /* to shutup gcc */
1144 ir_node *scnd_val = NULL; /* to shutup gcc */
1146 if (!get_opt_normalize()) return n;
1148 n_preds = get_Phi_n_preds(n);
1150 block = get_nodes_block(n);
1151 /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
1152 assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
1153 if ((is_Block_dead(block)) || /* Control dead */
1154 (block == current_ir_graph->start_block)) /* There should be no Phi nodes */
1155 return new_Bad(); /* in the Start Block. */
1157 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1159 /* If the Block has a Bad pred, we also have one. */
1160 for (i = 0; i < n_preds; ++i)
1161 if (is_Bad(get_Block_cfgpred(block, i)))
1162 set_Phi_pred(n, i, new_Bad());
1164 /* Find first non-self-referencing input */
1165 for (i = 0; i < n_preds; ++i) {
1166 first_val = get_Phi_pred(n, i);
1167 if ( (first_val != n) /* not self pointer */
1169 && (! is_Bad(first_val))
1171 ) { /* value not dead */
1172 break; /* then found first value. */
1177 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1183 /* follow_Id () for rest of inputs, determine if any of these
1184 are non-self-referencing */
1185 while (++i < n_preds) {
1186 scnd_val = get_Phi_pred(n, i);
1187 if ( (scnd_val != n)
1188 && (scnd_val != first_val)
1190 && (! is_Bad(scnd_val))
1198 /* Fold, if no multiple distinct non-self-referencing inputs */
1200 DBG_OPT_PHI(oldn, n);
1202 /* skip the remaining Ids (done in get_Phi_pred). */
1203 /* superfluous, since we walk all to propagate Block's Bads.
1204 while (++i < n_preds) get_Phi_pred(n, i); */
1210 * optimize Proj(Tuple) and gigo() for ProjX in Bad block
1212 static ir_node *equivalent_node_Proj(ir_node *n)
1216 ir_node *a = get_Proj_pred(n);
1218 if ( get_irn_op(a) == op_Tuple) {
1219 /* Remove the Tuple/Proj combination. */
1220 if ( get_Proj_proj(n) <= get_Tuple_n_preds(a) ) {
1221 n = get_Tuple_pred(a, get_Proj_proj(n));
1222 DBG_OPT_TUPLE(oldn, a, n);
1224 assert(0); /* This should not happen! */
1227 } else if (get_irn_mode(n) == mode_X) {
1228 if (is_Block_dead(get_nodes_block(skip_Proj(n)))) {
1229 /* Remove dead control flow -- early gigo(). */
1239 static ir_node *equivalent_node_Id(ir_node *n)
1245 } while (get_irn_op(n) == op_Id);
1247 DBG_OPT_ID(oldn, n);
1254 static ir_node *equivalent_node_Mux(ir_node *n)
1256 ir_node *oldn = n, *sel = get_Mux_sel(n);
1257 tarval *ts = value_of(sel);
1259 /* Mux(true, f, t) == t */
1260 if (ts == tarval_b_true) {
1261 n = get_Mux_true(n);
1262 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1264 /* Mux(false, f, t) == f */
1265 else if (ts == tarval_b_false) {
1266 n = get_Mux_false(n);
1267 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1269 /* Mux(v, x, x) == x */
1270 else if (get_Mux_false(n) == get_Mux_true(n)) {
1271 n = get_Mux_true(n);
1272 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1274 else if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(get_irn_mode(n))) {
1275 ir_node *cmp = get_Proj_pred(sel);
1276 long proj_nr = get_Proj_proj(sel);
1277 ir_node *b = get_Mux_false(n);
1278 ir_node *a = get_Mux_true(n);
1281 * Note: normalization puts the constant on the right site,
1282 * so we check only one case.
1284 * Note further that these optimization work even for floating point
1285 * with NaN's because -NaN == NaN.
1286 * However, if +0 and -0 is handled differently, we cannot use the first one.
1288 if (get_irn_op(cmp) == op_Cmp && get_Cmp_left(cmp) == a) {
1289 if (classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
1290 /* Mux(a CMP 0, X, a) */
1291 if (get_irn_op(b) == op_Minus && get_Minus_op(b) == a) {
1292 /* Mux(a CMP 0, -a, a) */
1293 if (proj_nr == pn_Cmp_Eq) {
1294 /* Mux(a == 0, -a, a) ==> -a */
1296 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1298 else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1299 /* Mux(a != 0, -a, a) ==> a */
1301 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1304 else if (classify_Const(b) == CNST_NULL) {
1305 /* Mux(a CMP 0, 0, a) */
1306 if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1307 /* Mux(a != 0, 0, a) ==> a */
1309 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1311 else if (proj_nr == pn_Cmp_Eq) {
1312 /* Mux(a == 0, 0, a) ==> 0 */
1314 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1325 * Optimize -a CMP -b into b CMP a.
1326 * This works only for for modes where unary Minus
1328 * Note that two-complement integers can Overflow
1329 * so it will NOT work.
1331 static ir_node *equivalent_node_Cmp(ir_node *n)
1333 ir_node *left = get_Cmp_left(n);
1334 ir_node *right = get_Cmp_right(n);
1336 if (get_irn_op(left) == op_Minus && get_irn_op(right) == op_Minus &&
1337 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1338 left = get_Minus_op(left);
1339 right = get_Minus_op(right);
1340 set_Cmp_left(n, right);
1341 set_Cmp_right(n, left);
1347 * Remove Confirm nodes if setting is on.
1349 static ir_node *equivalent_node_Confirm(ir_node *n)
1351 if (get_Confirm_cmp(n) == pn_Cmp_Eq) {
1352 ir_node *bound = get_Confirm_bound(n);
1353 ir_op *op = get_irn_op(bound);
1356 * Optimize a rare case:
1357 * Confirm(x, '=', Const) ==> Const
1359 if (op == op_Const || op == op_SymConst)
1362 return get_opt_remove_Confirm() ? get_Confirm_value(n) : n;
1366 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1367 * perform no actual computation, as, e.g., the Id nodes. It does not create
1368 * new nodes. It is therefore safe to free n if the node returned is not n.
1369 * If a node returns a Tuple we can not just skip it. If the size of the
1370 * in array fits, we transform n into a tuple (e.g., Div).
1373 equivalent_node(ir_node *n)
1375 if (n->op->equivalent_node)
1376 return n->op->equivalent_node(n);
1381 * set the default equivalent node operation
1383 static ir_op *firm_set_default_equivalent_node(ir_op *op)
1387 op->equivalent_node = equivalent_node_##a; \
1416 op->equivalent_node = NULL;
1424 * Do node specific optimizations of nodes predecessors.
1427 optimize_preds(ir_node *n) {
1428 ir_node *a = NULL, *b = NULL;
1430 /* get the operands we will work on for simple cases. */
1432 a = get_binop_left(n);
1433 b = get_binop_right(n);
1434 } else if (is_unop(n)) {
1438 switch (get_irn_opcode(n)) {
1441 /* We don't want Cast as input to Cmp. */
1442 if (get_irn_op(a) == op_Cast) {
1446 if (get_irn_op(b) == op_Cast) {
1448 set_Cmp_right(n, b);
1457 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1458 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1459 * If possible, remove the Conv's.
1461 static ir_node *transform_node_AddSub(ir_node *n)
1463 ir_mode *mode = get_irn_mode(n);
1465 if (mode_is_reference(mode)) {
1466 ir_node *left = get_binop_left(n);
1467 ir_node *right = get_binop_right(n);
1468 int ref_bits = get_mode_size_bits(mode);
1470 if (get_irn_op(left) == op_Conv) {
1471 ir_mode *mode = get_irn_mode(left);
1472 int bits = get_mode_size_bits(mode);
1474 if (ref_bits == bits &&
1475 mode_is_int(mode) &&
1476 get_mode_arithmetic(mode) == irma_twos_complement) {
1477 ir_node *pre = get_Conv_op(left);
1478 ir_mode *pre_mode = get_irn_mode(pre);
1480 if (mode_is_int(pre_mode) &&
1481 get_mode_size_bits(pre_mode) == bits &&
1482 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1483 /* ok, this conv just changes to sign, moreover the calculation
1484 * is done with same number of bits as our address mode, so
1485 * we can ignore the conv as address calculation can be viewed
1486 * as either signed or unsigned
1488 set_binop_left(n, pre);
1493 if (get_irn_op(right) == op_Conv) {
1494 ir_mode *mode = get_irn_mode(right);
1495 int bits = get_mode_size_bits(mode);
1497 if (ref_bits == bits &&
1498 mode_is_int(mode) &&
1499 get_mode_arithmetic(mode) == irma_twos_complement) {
1500 ir_node *pre = get_Conv_op(right);
1501 ir_mode *pre_mode = get_irn_mode(pre);
1503 if (mode_is_int(pre_mode) &&
1504 get_mode_size_bits(pre_mode) == bits &&
1505 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1506 /* ok, this conv just changes to sign, moreover the calculation
1507 * is done with same number of bits as our address mode, so
1508 * we can ignore the conv as address calculation can be viewed
1509 * as either signed or unsigned
1511 set_binop_right(n, pre);
1520 * Do the AddSub optimization, then Transform Add(a,a) into Mul(a, 2)
1521 * if the mode is integer or float.
1522 * Transform Add(a,-b) into Sub(a,b).
1523 * Reassociation might fold this further.
1525 static ir_node *transform_node_Add(ir_node *n)
1530 n = transform_node_AddSub(n);
1532 mode = get_irn_mode(n);
1533 if (mode_is_num(mode)) {
1534 ir_node *a = get_Add_left(n);
1536 if (a == get_Add_right(n)) {
1537 ir_node *block = get_irn_n(n, -1);
1540 get_irn_dbg_info(n),
1544 new_r_Const_long(current_ir_graph, block, mode, 2),
1546 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
1549 ir_node *b = get_Add_right(n);
1551 if (get_irn_op(a) == op_Minus) {
1553 get_irn_dbg_info(n),
1559 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1561 else if (get_irn_op(b) == op_Minus) {
1563 get_irn_dbg_info(n),
1569 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1577 * Do the AddSub optimization, then Transform Sub(0,a) into Minus(a).
1579 static ir_node *transform_node_Sub(ir_node *n)
1584 n = transform_node_AddSub(n);
1586 mode = get_irn_mode(n);
1587 if (mode_is_num(mode) && (classify_Const(get_Sub_left(n)) == CNST_NULL)) {
1589 get_irn_dbg_info(n),
1594 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
1601 * Transform Mul(a,-1) into -a.
1602 * Do architecture dependent optimizations on Mul nodes
1604 static ir_node *transform_node_Mul(ir_node *n) {
1606 ir_mode *mode = get_irn_mode(n);
1608 if (mode_is_signed(mode)) {
1610 ir_node *a = get_Mul_left(n);
1611 ir_node *b = get_Mul_right(n);
1613 if (value_of(a) == get_mode_minus_one(mode))
1615 else if (value_of(b) == get_mode_minus_one(mode))
1618 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
1619 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
1623 return arch_dep_replace_mul_with_shifts(n);
1627 * transform a Div Node
1629 static ir_node *transform_node_Div(ir_node *n)
1631 tarval *tv = value_of(n);
1634 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
1636 if (tv != tarval_bad) {
1637 value = new_Const(get_tarval_mode(tv), tv);
1639 DBG_OPT_CSTEVAL(n, value);
1641 else /* Try architecture dependent optimization */
1642 value = arch_dep_replace_div_by_const(n);
1645 /* Turn Div into a tuple (mem, bad, value) */
1646 ir_node *mem = get_Div_mem(n);
1648 turn_into_tuple(n, 3);
1649 set_Tuple_pred(n, pn_Div_M, mem);
1650 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
1651 set_Tuple_pred(n, pn_Div_res, value);
1657 * transform a Mod node
1659 static ir_node *transform_node_Mod(ir_node *n)
1661 tarval *tv = value_of(n);
1664 /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */
1666 if (tv != tarval_bad) {
1667 value = new_Const(get_tarval_mode(tv), tv);
1669 DBG_OPT_CSTEVAL(n, value);
1671 else /* Try architecture dependent optimization */
1672 value = arch_dep_replace_mod_by_const(n);
1675 /* Turn Mod into a tuple (mem, bad, value) */
1676 ir_node *mem = get_Mod_mem(n);
1678 turn_into_tuple(n, 3);
1679 set_Tuple_pred(n, pn_Mod_M, mem);
1680 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
1681 set_Tuple_pred(n, pn_Mod_res, value);
1687 * transform a DivMod node
1689 static ir_node *transform_node_DivMod(ir_node *n)
1693 ir_node *a = get_DivMod_left(n);
1694 ir_node *b = get_DivMod_right(n);
1695 ir_mode *mode = get_irn_mode(a);
1696 tarval *ta = value_of(a);
1697 tarval *tb = value_of(b);
1699 if (!(mode_is_int(mode) && mode_is_int(get_irn_mode(b))))
1702 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
1704 if (tb != tarval_bad) {
1705 if (tb == get_mode_one(get_tarval_mode(tb))) {
1706 b = new_Const (mode, get_mode_null(mode));
1709 DBG_OPT_CSTEVAL(n, b);
1711 else if (ta != tarval_bad) {
1712 tarval *resa, *resb;
1713 resa = tarval_div (ta, tb);
1714 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
1715 Jmp for X result!? */
1716 resb = tarval_mod (ta, tb);
1717 if (resb == tarval_bad) return n; /* Causes exception! */
1718 a = new_Const (mode, resa);
1719 b = new_Const (mode, resb);
1722 DBG_OPT_CSTEVAL(n, a);
1723 DBG_OPT_CSTEVAL(n, b);
1725 else { /* Try architecture dependent optimization */
1726 arch_dep_replace_divmod_by_const(&a, &b, n);
1727 evaluated = a != NULL;
1729 } else if (ta == get_mode_null(mode)) {
1730 /* 0 / non-Const = 0 */
1735 if (evaluated) { /* replace by tuple */
1736 ir_node *mem = get_DivMod_mem(n);
1737 turn_into_tuple(n, 4);
1738 set_Tuple_pred(n, pn_DivMod_M, mem);
1739 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1740 set_Tuple_pred(n, pn_DivMod_res_div, a);
1741 set_Tuple_pred(n, pn_DivMod_res_mod, b);
1748 * Optimize Abs(x) into x if x is Confirmed >= 0
1749 * Optimize Abs(x) into -x if x is Confirmed <= 0
1751 static ir_node *transform_node_Abs(ir_node *n)
1754 ir_node *a = get_Abs_op(n);
1755 value_classify sign = classify_value_sign(a);
1757 if (sign == VALUE_NEGATIVE) {
1758 ir_mode *mode = get_irn_mode(n);
1761 * We can replace the Abs by -x here.
1762 * We even could add a new Confirm here.
1764 * Note that -x would create a new node, so we could
1765 * not run it in the equivalent_node() context.
1767 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
1768 get_irn_n(n, -1), a, mode);
1770 DBG_OPT_CONFIRM(oldn, n);
1772 else if (sign == VALUE_POSITIVE) {
1773 /* n is positive, Abs is not needed */
1776 DBG_OPT_CONFIRM(oldn, n);
1783 * transform a Cond node
1785 static ir_node *transform_node_Cond(ir_node *n)
1787 /* Replace the Cond by a Jmp if it branches on a constant
1790 ir_node *a = get_Cond_selector(n);
1791 tarval *ta = value_of(a);
1793 /* we need block info which is not available in floating irgs */
1794 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
1797 if ((ta != tarval_bad) &&
1798 (get_irn_mode(a) == mode_b) &&
1799 (get_opt_unreachable_code())) {
1800 /* It's a boolean Cond, branching on a boolean constant.
1801 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
1802 jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n));
1803 turn_into_tuple(n, 2);
1804 if (ta == tarval_b_true) {
1805 set_Tuple_pred(n, pn_Cond_false, new_Bad());
1806 set_Tuple_pred(n, pn_Cond_true, jmp);
1808 set_Tuple_pred(n, pn_Cond_false, jmp);
1809 set_Tuple_pred(n, pn_Cond_true, new_Bad());
1811 /* We might generate an endless loop, so keep it alive. */
1812 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
1813 } else if ((ta != tarval_bad) &&
1814 (get_irn_mode(a) == mode_Iu) &&
1815 (get_Cond_kind(n) == dense) &&
1816 (get_opt_unreachable_code())) {
1817 /* I don't want to allow Tuples smaller than the biggest Proj.
1818 Also this tuple might get really big...
1819 I generate the Jmp here, and remember it in link. Link is used
1820 when optimizing Proj. */
1821 set_irn_link(n, new_r_Jmp(current_ir_graph, get_nodes_block(n)));
1822 /* We might generate an endless loop, so keep it alive. */
1823 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
1824 } else if ((get_irn_op(a) == op_Eor)
1825 && (get_irn_mode(a) == mode_b)
1826 && (classify_tarval(value_of(get_Eor_right(a))) == TV_CLASSIFY_ONE)) {
1827 /* The Eor is a negate. Generate a new Cond without the negate,
1828 simulate the negate by exchanging the results. */
1829 set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_block(n),
1831 } else if ((get_irn_op(a) == op_Not)
1832 && (get_irn_mode(a) == mode_b)) {
1833 /* A Not before the Cond. Generate a new Cond without the Not,
1834 simulate the Not by exchanging the results. */
1835 set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_block(n),
1844 static ir_node *transform_node_Eor(ir_node *n)
1847 ir_node *a = get_Eor_left(n);
1848 ir_node *b = get_Eor_right(n);
1849 ir_mode *mode = get_irn_mode(n);
1853 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
1854 mode, get_mode_null(mode));
1855 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
1857 else if ((mode == mode_b)
1858 && (get_irn_op(a) == op_Proj)
1859 && (get_irn_mode(a) == mode_b)
1860 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)
1861 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
1862 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
1863 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
1864 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
1866 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
1868 else if ((mode == mode_b)
1869 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) {
1870 /* The Eor is a Not. Replace it by a Not. */
1871 /* ????!!!Extend to bitfield 1111111. */
1872 n = new_r_Not(current_ir_graph, get_irn_n(n, -1), a, mode_b);
1874 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
1881 * Transform a boolean Not.
1883 static ir_node *transform_node_Not(ir_node *n)
1886 ir_node *a = get_Not_op(n);
1888 if ( (get_irn_mode(n) == mode_b)
1889 && (get_irn_op(a) == op_Proj)
1890 && (get_irn_mode(a) == mode_b)
1891 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
1892 /* We negate a Cmp. The Cmp has the negated result anyways! */
1893 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
1894 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
1895 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
1902 * Transform a Cast_type(Const) into a new Const_type
1904 static ir_node *transform_node_Cast(ir_node *n) {
1906 ir_node *pred = get_Cast_op(n);
1907 type *tp = get_irn_type(n);
1909 if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) {
1910 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
1911 get_Const_tarval(pred), tp);
1912 DBG_OPT_CSTEVAL(oldn, n);
1913 } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) {
1914 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred),
1915 get_SymConst_kind(pred), tp);
1916 DBG_OPT_CSTEVAL(oldn, n);
1923 * Transform a Proj(Div) with a non-zero value.
1924 * Removes the exceptions and routes the memory to the NoMem node.
1926 static ir_node *transform_node_Proj_Div(ir_node *proj)
1928 ir_node *n = get_Proj_pred(proj);
1929 ir_node *b = get_Div_right(n);
1932 if (value_not_zero(b)) {
1933 /* div(x, y) && y != 0 */
1934 proj_nr = get_Proj_proj(proj);
1936 /* this node may float */
1937 set_irn_pinned(n, op_pin_state_floats);
1939 if (proj_nr == pn_Div_X_except) {
1940 /* we found an exception handler, remove it */
1943 /* the memory Proj can be removed */
1944 ir_node *res = get_Div_mem(n);
1945 set_Div_mem(n, get_irg_no_mem(current_ir_graph));
1946 if (proj_nr == pn_Div_M)
1954 * Transform a Proj(Mod) with a non-zero value.
1955 * Removes the exceptions and routes the memory to the NoMem node.
1957 static ir_node *transform_node_Proj_Mod(ir_node *proj)
1959 ir_node *n = get_Proj_pred(proj);
1960 ir_node *b = get_Mod_right(n);
1963 if (value_not_zero(b)) {
1964 /* mod(x, y) && y != 0 */
1965 proj_nr = get_Proj_proj(proj);
1967 /* this node may float */
1968 set_irn_pinned(n, op_pin_state_floats);
1970 if (proj_nr == pn_Mod_X_except) {
1971 /* we found an exception handler, remove it */
1974 /* the memory Proj can be removed */
1975 ir_node *res = get_Mod_mem(n);
1976 set_Mod_mem(n, get_irg_no_mem(current_ir_graph));
1977 if (proj_nr == pn_Mod_M)
1985 * Transform a Proj(DivMod) with a non-zero value.
1986 * Removes the exceptions and routes the memory to the NoMem node.
1988 static ir_node *transform_node_Proj_DivMod(ir_node *proj)
1990 ir_node *n = get_Proj_pred(proj);
1991 ir_node *b = get_DivMod_right(n);
1994 if (value_not_zero(b)) {
1995 /* DivMod(x, y) && y != 0 */
1996 proj_nr = get_Proj_proj(proj);
1998 /* this node may float */
1999 set_irn_pinned(n, op_pin_state_floats);
2001 if (proj_nr == pn_DivMod_X_except) {
2002 /* we found an exception handler, remove it */
2006 /* the memory Proj can be removed */
2007 ir_node *res = get_DivMod_mem(n);
2008 set_DivMod_mem(n, get_irg_no_mem(current_ir_graph));
2009 if (proj_nr == pn_DivMod_M)
2017 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
2019 static ir_node *transform_node_Proj_Cond(ir_node *proj)
2021 if (get_opt_unreachable_code()) {
2022 ir_node *n = get_Proj_pred(proj);
2023 ir_node *b = get_Cond_selector(n);
2024 tarval *tb = value_of(b);
2026 if (tb != tarval_bad && mode_is_int(get_tarval_mode(tb))) {
2027 /* we have a constant switch */
2028 long num = get_Proj_proj(proj);
2030 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
2031 if (get_tarval_long(tb) == num) {
2032 /* Do NOT create a jump here, or we will have 2 control flow ops
2033 * in a block. This case is optimized away in optimize_cf(). */
2037 /* this case will NEVER be taken, kill it */
2047 * Normalizes and optimizes Cmp nodes.
2049 static ir_node *transform_node_Proj_Cmp(ir_node *proj)
2051 if (get_opt_reassociation()) {
2052 ir_node *n = get_Proj_pred(proj);
2053 ir_node *left = get_Cmp_left(n);
2054 ir_node *right = get_Cmp_right(n);
2058 ir_mode *mode = NULL;
2059 long proj_nr = get_Proj_proj(proj);
2062 * First step: normalize the compare op
2063 * by placing the constant on the right site
2064 * or moving the lower address node to the left.
2065 * We ignore the case that both are constants
2066 * this case should be optimized away.
2068 if (get_irn_op(right) == op_Const)
2070 else if (get_irn_op(left) == op_Const) {
2075 proj_nr = get_inversed_pnc(proj_nr);
2078 else if (left > right) {
2084 proj_nr = get_inversed_pnc(proj_nr);
2089 * Second step: Try to reduce the magnitude
2090 * of a constant. This may help to generate better code
2091 * later and may help to normalize more compares.
2092 * Of course this is only possible for integer values.
2095 mode = get_irn_mode(c);
2096 tv = get_Const_tarval(c);
2098 if (tv != tarval_bad) {
2099 /* the following optimization is possible on modes without Overflow
2100 * on Unary Minus or on == and !=:
2101 * -a CMP c ==> a swap(CMP) -c
2103 * Beware: for two-complement Overflow may occur, so only == and != can
2104 * be optimized, see this:
2105 * -MININT < 0 =/=> MININT > 0 !!!
2107 if (get_opt_constant_folding() && get_irn_op(left) == op_Minus &&
2108 (!mode_overflow_on_unary_Minus(mode) ||
2109 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
2110 left = get_Minus_op(left);
2111 tv = tarval_sub(get_mode_null(mode), tv);
2113 proj_nr = get_inversed_pnc(proj_nr);
2117 /* for integer modes, we have more */
2118 if (mode_is_int(mode)) {
2119 /* Ne includes Unordered which is not possible on integers.
2120 * However, frontends often use this wrong, so fix it here */
2121 if (proj_nr == pn_Cmp_Ne) {
2122 proj_nr = pn_Cmp_Lg;
2123 set_Proj_proj(proj, proj_nr);
2126 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
2127 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
2128 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
2129 tv = tarval_sub(tv, get_mode_one(mode));
2131 proj_nr ^= pn_Cmp_Eq;
2134 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
2135 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
2136 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
2137 tv = tarval_add(tv, get_mode_one(mode));
2139 proj_nr ^= pn_Cmp_Eq;
2143 /* the following reassociations work only for == and != */
2144 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
2146 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
2147 if (classify_tarval(tv) == TV_CLASSIFY_NULL && get_irn_op(left) == op_Sub) {
2148 right = get_Sub_right(left);
2149 left = get_Sub_left(left);
2151 tv = value_of(right);
2155 if (tv != tarval_bad) {
2156 ir_op *op = get_irn_op(left);
2158 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
2160 ir_node *c1 = get_Sub_right(left);
2161 tarval *tv2 = value_of(c1);
2163 if (tv2 != tarval_bad) {
2164 tv2 = tarval_add(tv, value_of(c1));
2166 if (tv2 != tarval_bad) {
2167 left = get_Sub_left(left);
2173 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
2174 else if (op == op_Add) {
2175 ir_node *a_l = get_Add_left(left);
2176 ir_node *a_r = get_Add_right(left);
2180 if (get_irn_op(a_l) == op_Const) {
2182 tv2 = value_of(a_l);
2186 tv2 = value_of(a_r);
2189 if (tv2 != tarval_bad) {
2190 tv2 = tarval_sub(tv, tv2);
2192 if (tv2 != tarval_bad) {
2199 /* -a == c ==> a == -c, -a != c ==> a != -c */
2200 else if (op == op_Minus) {
2201 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
2203 if (tv2 != tarval_bad) {
2204 left = get_Minus_op(left);
2211 /* the following reassociations work only for <= */
2212 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2213 if (tv != tarval_bad) {
2214 ir_op *op = get_irn_op(left);
2216 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
2226 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
2228 if (changed & 2) /* need a new Const */
2229 right = new_Const(mode, tv);
2231 /* create a new compare */
2232 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block,
2235 set_Proj_pred(proj, n);
2236 set_Proj_proj(proj, proj_nr);
2243 * Does all optimizations on nodes that must be done on it's Proj's
2244 * because of creating new nodes.
2246 static ir_node *transform_node_Proj(ir_node *proj)
2248 ir_node *n = get_Proj_pred(proj);
2250 switch (get_irn_opcode(n)) {
2252 return transform_node_Proj_Div(proj);
2255 return transform_node_Proj_Mod(proj);
2258 return transform_node_Proj_DivMod(proj);
2261 return transform_node_Proj_Cond(proj);
2264 return transform_node_Proj_Cmp(proj);
2267 /* should not happen, but if it does will be optimized away */
2268 return equivalent_node_Proj(proj);
2277 * returns the operands of a commutative bin-op, if one operand is
2278 * a const, it is returned as the second one.
2280 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c)
2282 ir_node *op_a = get_binop_left(binop);
2283 ir_node *op_b = get_binop_right(binop);
2285 assert(is_op_commutative(get_irn_op(binop)));
2287 if (get_irn_op(op_a) == op_Const) {
2298 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
2299 * Such pattern may arise in bitfield stores.
2301 * value c4 value c4 & c2
2302 * AND c3 AND c1 | c3
2307 static ir_node *transform_node_Or_bf_store(ir_node *or)
2311 ir_node *and_l, *c3;
2312 ir_node *value, *c4;
2313 ir_node *new_and, *new_const, *block;
2314 ir_mode *mode = get_irn_mode(or);
2316 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
2318 get_comm_Binop_Ops(or, &and, &c1);
2319 if ((get_irn_op(c1) != op_Const) || (get_irn_op(and) != op_And))
2322 get_comm_Binop_Ops(and, &or_l, &c2);
2323 if ((get_irn_op(c2) != op_Const) || (get_irn_op(or_l) != op_Or))
2326 get_comm_Binop_Ops(or_l, &and_l, &c3);
2327 if ((get_irn_op(c3) != op_Const) || (get_irn_op(and_l) != op_And))
2330 get_comm_Binop_Ops(and_l, &value, &c4);
2331 if (get_irn_op(c4) != op_Const)
2334 /* ok, found the pattern, check for conditions */
2335 assert(mode == get_irn_mode(and));
2336 assert(mode == get_irn_mode(or_l));
2337 assert(mode == get_irn_mode(and_l));
2339 tv1 = get_Const_tarval(c1);
2340 tv2 = get_Const_tarval(c2);
2341 tv3 = get_Const_tarval(c3);
2342 tv4 = get_Const_tarval(c4);
2344 tv = tarval_or(tv4, tv2);
2345 if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) {
2346 /* have at least one 0 at the same bit position */
2350 n_tv4 = tarval_not(tv4);
2351 if (tv3 != tarval_and(tv3, n_tv4)) {
2352 /* bit in the or_mask is outside the and_mask */
2356 n_tv2 = tarval_not(tv2);
2357 if (tv1 != tarval_and(tv1, n_tv2)) {
2358 /* bit in the or_mask is outside the and_mask */
2362 /* ok, all conditions met */
2363 block = get_irn_n(or, -1);
2365 new_and = new_r_And(current_ir_graph, block,
2366 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
2368 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
2370 set_Or_left(or, new_and);
2371 set_Or_right(or, new_const);
2373 /* check for more */
2374 return transform_node_Or_bf_store(or);
2378 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
2380 static ir_node *transform_node_Or_Rot(ir_node *or)
2382 ir_mode *mode = get_irn_mode(or);
2383 ir_node *shl, *shr, *block;
2384 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
2387 if (! mode_is_int(mode))
2390 shl = get_binop_left(or);
2391 shr = get_binop_right(or);
2393 if (get_irn_op(shl) == op_Shr) {
2394 if (get_irn_op(shr) != op_Shl)
2401 else if (get_irn_op(shl) != op_Shl)
2403 else if (get_irn_op(shr) != op_Shr)
2406 x = get_Shl_left(shl);
2407 if (x != get_Shr_left(shr))
2410 c1 = get_Shl_right(shl);
2411 c2 = get_Shr_right(shr);
2412 if (get_irn_op(c1) == op_Const && get_irn_op(c2) == op_Const) {
2413 tv1 = get_Const_tarval(c1);
2414 if (! tarval_is_long(tv1))
2417 tv2 = get_Const_tarval(c2);
2418 if (! tarval_is_long(tv2))
2421 if (get_tarval_long(tv1) + get_tarval_long(tv2)
2422 != get_mode_size_bits(mode))
2425 /* yet, condition met */
2426 block = get_irn_n(or, -1);
2428 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
2430 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
2433 else if (get_irn_op(c1) == op_Sub) {
2437 if (get_Sub_right(sub) != v)
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);
2454 /* a Rot right is not supported, so use a rot left */
2455 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
2457 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
2460 else if (get_irn_op(c2) == op_Sub) {
2464 c1 = get_Sub_left(sub);
2465 if (get_irn_op(c1) != op_Const)
2468 tv1 = get_Const_tarval(c1);
2469 if (! tarval_is_long(tv1))
2472 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
2475 /* yet, condition met */
2476 block = get_irn_n(or, -1);
2479 n = new_r_Rot(current_ir_graph, block, x, v, mode);
2481 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
2491 static ir_node *transform_node_Or(ir_node *or)
2493 or = transform_node_Or_bf_store(or);
2494 or = transform_node_Or_Rot(or);
2500 static ir_node *transform_node(ir_node *n);
2503 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl.
2505 * Should be moved to reassociation?
2507 static ir_node *transform_node_shift(ir_node *n)
2509 ir_node *left, *right;
2510 tarval *tv1, *tv2, *res;
2512 int modulo_shf, flag;
2514 left = get_binop_left(n);
2516 /* different operations */
2517 if (get_irn_op(left) != get_irn_op(n))
2520 right = get_binop_right(n);
2521 tv1 = value_of(right);
2522 if (tv1 == tarval_bad)
2525 tv2 = value_of(get_binop_right(left));
2526 if (tv2 == tarval_bad)
2529 res = tarval_add(tv1, tv2);
2531 /* beware: a simple replacement works only, if res < modulo shift */
2532 mode = get_irn_mode(n);
2536 modulo_shf = get_mode_modulo_shift(mode);
2537 if (modulo_shf > 0) {
2538 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
2540 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
2547 /* ok, we can replace it */
2548 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
2550 in[0] = get_binop_left(left);
2551 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
2553 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
2555 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
2557 return transform_node(irn);
2562 #define transform_node_Shr transform_node_shift
2563 #define transform_node_Shrs transform_node_shift
2564 #define transform_node_Shl transform_node_shift
2567 * Remove dead blocks and nodes in dead blocks
2568 * in keep alive list. We do not generate a new End node.
2570 static ir_node *transform_node_End(ir_node *n) {
2571 int i, n_keepalives = get_End_n_keepalives(n);
2573 for (i = 0; i < n_keepalives; ++i) {
2574 ir_node *ka = get_End_keepalive(n, i);
2576 if (is_Block_dead(ka)) {
2577 set_End_keepalive(n, i, new_Bad());
2580 else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka)))
2581 set_End_keepalive(n, i, new_Bad());
2587 * Optimize a Mux into some simpler cases.
2589 static ir_node *transform_node_Mux(ir_node *n)
2591 ir_node *oldn = n, *sel = get_Mux_sel(n);
2592 ir_mode *mode = get_irn_mode(n);
2594 if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(mode)) {
2595 ir_node *cmp = get_Proj_pred(sel);
2596 long proj_nr = get_Proj_proj(sel);
2597 ir_node *f = get_Mux_false(n);
2598 ir_node *t = get_Mux_true(n);
2600 if (get_irn_op(cmp) == op_Cmp && classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
2601 ir_node *block = get_irn_n(n, -1);
2604 * Note: normalization puts the constant on the right site,
2605 * so we check only one case.
2607 * Note further that these optimization work even for floating point
2608 * with NaN's because -NaN == NaN.
2609 * However, if +0 and -0 is handled differently, we cannot use the first one.
2611 if (get_irn_op(f) == op_Minus &&
2612 get_Minus_op(f) == t &&
2613 get_Cmp_left(cmp) == t) {
2615 if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
2616 /* Mux(a >=/> 0, -a, a) ==> Abs(a) */
2617 n = new_rd_Abs(get_irn_dbg_info(n),
2621 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2624 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2625 /* Mux(a <=/< 0, -a, a) ==> Minus(Abs(a)) */
2626 n = new_rd_Abs(get_irn_dbg_info(n),
2630 n = new_rd_Minus(get_irn_dbg_info(n),
2635 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2639 else if (get_irn_op(t) == op_Minus &&
2640 get_Minus_op(t) == f &&
2641 get_Cmp_left(cmp) == f) {
2643 if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2644 /* Mux(a <=/< 0, a, -a) ==> Abs(a) */
2645 n = new_rd_Abs(get_irn_dbg_info(n),
2649 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2652 else if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
2653 /* Mux(a >=/> 0, a, -a) ==> Minus(Abs(a)) */
2654 n = new_rd_Abs(get_irn_dbg_info(n),
2658 n = new_rd_Minus(get_irn_dbg_info(n),
2663 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2668 if (mode_is_int(mode) && mode_is_signed(mode) &&
2669 get_mode_arithmetic(mode) == irma_twos_complement) {
2670 ir_node *x = get_Cmp_left(cmp);
2672 /* the following optimization works only with signed integer two-complement mode */
2674 if (mode == get_irn_mode(x)) {
2676 * FIXME: this restriction is two rigid, as it would still
2677 * work if mode(x) = Hs and mode == Is, but at least it removes
2680 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Le) &&
2681 classify_Const(t) == CNST_ALL_ONE &&
2682 classify_Const(f) == CNST_NULL) {
2684 * Mux(x:T </<= 0, 0, -1) -> Shrs(x, sizeof_bits(T) - 1)
2688 n = new_rd_Shrs(get_irn_dbg_info(n),
2689 current_ir_graph, block, x,
2690 new_r_Const_long(current_ir_graph, block, mode_Iu,
2691 get_mode_size_bits(mode) - 1),
2693 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
2696 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) &&
2697 classify_Const(t) == CNST_ONE &&
2698 classify_Const(f) == CNST_NULL) {
2700 * Mux(x:T >/>= 0, 0, 1) -> Shr(-x, sizeof_bits(T) - 1)
2704 n = new_rd_Shr(get_irn_dbg_info(n),
2705 current_ir_graph, block,
2706 new_r_Minus(current_ir_graph, block, x, mode),
2707 new_r_Const_long(current_ir_graph, block, mode_Iu,
2708 get_mode_size_bits(mode) - 1),
2710 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
2717 return arch_transform_node_Mux(n);
2721 * Tries several [inplace] [optimizing] transformations and returns an
2722 * equivalent node. The difference to equivalent_node() is that these
2723 * transformations _do_ generate new nodes, and thus the old node must
2724 * not be freed even if the equivalent node isn't the old one.
2726 static ir_node *transform_node(ir_node *n)
2728 if (n->op->transform_node)
2729 n = n->op->transform_node(n);
2734 * set the default transform node operation
2736 static ir_op *firm_set_default_transform_node(ir_op *op)
2740 op->transform_node = transform_node_##a; \
2764 op->transform_node = NULL;
2772 /* **************** Common Subexpression Elimination **************** */
2774 /** The size of the hash table used, should estimate the number of nodes
2776 #define N_IR_NODES 512
2778 /** Compares the attributes of two Const nodes. */
2779 static int node_cmp_attr_Const(ir_node *a, ir_node *b)
2781 return (get_Const_tarval(a) != get_Const_tarval(b))
2782 || (get_Const_type(a) != get_Const_type(b));
2785 /** Compares the attributes of two Proj nodes. */
2786 static int node_cmp_attr_Proj(ir_node *a, ir_node *b)
2788 return get_irn_proj_attr (a) != get_irn_proj_attr (b);
2791 /** Compares the attributes of two Filter nodes. */
2792 static int node_cmp_attr_Filter(ir_node *a, ir_node *b)
2794 return get_Filter_proj(a) != get_Filter_proj(b);
2797 /** Compares the attributes of two Alloc nodes. */
2798 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b)
2800 return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where)
2801 || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type);
2804 /** Compares the attributes of two Free nodes. */
2805 static int node_cmp_attr_Free(ir_node *a, ir_node *b)
2807 return (get_irn_free_attr(a).where != get_irn_free_attr(b).where)
2808 || (get_irn_free_attr(a).type != get_irn_free_attr(b).type);
2811 /** Compares the attributes of two SymConst nodes. */
2812 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b)
2814 return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num)
2815 || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p)
2816 || (get_irn_symconst_attr(a).tp != get_irn_symconst_attr(b).tp);
2819 /** Compares the attributes of two Call nodes. */
2820 static int node_cmp_attr_Call(ir_node *a, ir_node *b)
2822 return (get_irn_call_attr(a) != get_irn_call_attr(b));
2825 /** Compares the attributes of two Sel nodes. */
2826 static int node_cmp_attr_Sel(ir_node *a, ir_node *b)
2828 return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind)
2829 || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name)
2830 || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner)
2831 || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name)
2832 || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type);
2835 /** Compares the attributes of two Phi nodes. */
2836 static int node_cmp_attr_Phi(ir_node *a, ir_node *b)
2838 return get_irn_phi_attr (a) != get_irn_phi_attr (b);
2841 /** Compares the attributes of two Cast nodes. */
2842 static int node_cmp_attr_Cast(ir_node *a, ir_node *b)
2844 return get_Cast_type(a) != get_Cast_type(b);
2847 /** Compares the attributes of two Load nodes. */
2848 static int node_cmp_attr_Load(ir_node *a, ir_node *b)
2850 if (get_Load_volatility(a) == volatility_is_volatile ||
2851 get_Load_volatility(b) == volatility_is_volatile)
2852 /* NEVER do CSE on volatile Loads */
2855 return get_Load_mode(a) != get_Load_mode(b);
2858 /** Compares the attributes of two Store nodes. */
2859 static int node_cmp_attr_Store(ir_node *a, ir_node *b)
2861 /* NEVER do CSE on volatile Stores */
2862 return (get_Store_volatility(a) == volatility_is_volatile ||
2863 get_Store_volatility(b) == volatility_is_volatile);
2866 /** Compares the attributes of two Confirm nodes. */
2867 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b)
2869 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
2873 * set the default node attribute compare operation
2875 static ir_op *firm_set_default_node_cmp_attr(ir_op *op)
2879 op->node_cmp_attr = node_cmp_attr_##a; \
2897 op->node_cmp_attr = NULL;
2905 * Compare function for two nodes in the hash table. Gets two
2906 * nodes as parameters. Returns 0 if the nodes are a cse.
2909 vt_cmp (const void *elt, const void *key)
2917 if (a == b) return 0;
2919 if ((get_irn_op(a) != get_irn_op(b)) ||
2920 (get_irn_mode(a) != get_irn_mode(b))) return 1;
2922 /* compare if a's in and b's in are of equal length */
2923 irn_arity_a = get_irn_intra_arity (a);
2924 if (irn_arity_a != get_irn_intra_arity(b))
2927 /* for block-local cse and op_pin_state_pinned nodes: */
2928 if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) {
2929 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
2933 /* compare a->in[0..ins] with b->in[0..ins] */
2934 for (i = 0; i < irn_arity_a; i++)
2935 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
2939 * here, we already now that the nodes are identical except their
2942 if (a->op->node_cmp_attr)
2943 return a->op->node_cmp_attr(a, b);
2949 * Calculate a hash value of a node.
2952 ir_node_hash (ir_node *node)
2957 if (node->op == op_Const) {
2958 /* special value for const, as they only differ in their tarval. */
2959 h = HASH_PTR(node->attr.con.tv);
2960 h = 9*h + HASH_PTR(get_irn_mode(node));
2961 } else if (node->op == op_SymConst) {
2962 /* special value for const, as they only differ in their symbol. */
2963 h = HASH_PTR(node->attr.i.sym.type_p);
2964 h = 9*h + HASH_PTR(get_irn_mode(node));
2967 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
2968 h = irn_arity = get_irn_intra_arity(node);
2970 /* consider all in nodes... except the block if not a control flow. */
2971 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
2972 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
2976 h = 9*h + HASH_PTR(get_irn_mode(node));
2978 h = 9*h + HASH_PTR(get_irn_op(node));
2985 new_identities(void) {
2986 return new_pset(vt_cmp, N_IR_NODES);
2990 del_identities(pset *value_table) {
2991 del_pset(value_table);
2995 * Return the canonical node computing the same value as n.
2996 * Looks up the node in a hash table.
2998 * For Const nodes this is performed in the constructor, too. Const
2999 * nodes are extremely time critical because of their frequent use in
3000 * constant string arrays.
3002 static INLINE ir_node *
3003 identify (pset *value_table, ir_node *n)
3007 if (!value_table) return n;
3009 if (get_opt_reassociation()) {
3010 if (is_op_commutative(get_irn_op(n))) {
3011 ir_node *l = get_binop_left(n);
3012 ir_node *r = get_binop_right(n);
3014 /* for commutative operators perform a OP b == b OP a */
3016 set_binop_left(n, r);
3017 set_binop_right(n, l);
3022 o = pset_find (value_table, n, ir_node_hash (n));
3031 * During construction we set the op_pin_state_pinned flag in the graph right when the
3032 * optimization is performed. The flag turning on procedure global cse could
3033 * be changed between two allocations. This way we are safe.
3035 static INLINE ir_node *
3036 identify_cons (pset *value_table, ir_node *n) {
3039 n = identify(value_table, n);
3040 if (get_irn_n(old, -1) != get_irn_n(n, -1))
3041 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3046 * Return the canonical node computing the same value as n.
3047 * Looks up the node in a hash table, enters it in the table
3048 * if it isn't there yet.
3051 identify_remember (pset *value_table, ir_node *n)
3055 if (!value_table) return n;
3057 if (get_opt_reassociation()) {
3058 if (is_op_commutative(get_irn_op(n))) {
3059 ir_node *l = get_binop_left(n);
3060 ir_node *r = get_binop_right(n);
3062 /* for commutative operators perform a OP b == b OP a */
3064 set_binop_left(n, r);
3065 set_binop_right(n, l);
3070 /* lookup or insert in hash table with given hash key. */
3071 o = pset_insert (value_table, n, ir_node_hash (n));
3081 add_identities (pset *value_table, ir_node *node) {
3082 if (get_opt_cse() && (get_irn_opcode(node) != iro_Block))
3083 identify_remember (value_table, node);
3087 * garbage in, garbage out. If a node has a dead input, i.e., the
3088 * Bad node is input to the node, return the Bad node.
3090 static INLINE ir_node *
3091 gigo (ir_node *node)
3094 ir_op *op = get_irn_op(node);
3096 /* remove garbage blocks by looking at control flow that leaves the block
3097 and replacing the control flow by Bad. */
3098 if (get_irn_mode(node) == mode_X) {
3099 ir_node *block = get_nodes_block(skip_Proj(node));
3101 /* Don't optimize nodes in immature blocks. */
3102 if (!get_Block_matured(block)) return node;
3103 /* Don't optimize End, may have Bads. */
3104 if (op == op_End) return node;
3106 if (is_Block(block)) {
3107 irn_arity = get_irn_arity(block);
3108 for (i = 0; i < irn_arity; i++) {
3109 if (!is_Bad(get_irn_n(block, i))) break;
3111 if (i == irn_arity) return new_Bad();
3115 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
3116 blocks predecessors is dead. */
3117 if ( op != op_Block && op != op_Phi && op != op_Tuple) {
3118 irn_arity = get_irn_arity(node);
3121 * Beware: we can only read the block of a non-floating node.
3123 if (is_irn_pinned_in_irg(node) &&
3124 is_Block_dead(get_nodes_block(node)))
3127 for (i = 0; i < irn_arity; i++) {
3128 if (is_Bad(get_irn_n(node, i))) {
3134 /* With this code we violate the agreement that local_optimize
3135 only leaves Bads in Block, Phi and Tuple nodes. */
3136 /* If Block has only Bads as predecessors it's garbage. */
3137 /* If Phi has only Bads as predecessors it's garbage. */
3138 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
3139 irn_arity = get_irn_arity(node);
3140 for (i = 0; i < irn_arity; i++) {
3141 if (!is_Bad(get_irn_n(node, i))) break;
3143 if (i == irn_arity) node = new_Bad();
3151 * These optimizations deallocate nodes from the obstack.
3152 * It can only be called if it is guaranteed that no other nodes
3153 * reference this one, i.e., right after construction of a node.
3156 optimize_node(ir_node *n)
3160 opcode iro = get_irn_opcode(n);
3162 /* Always optimize Phi nodes: part of the construction. */
3163 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
3165 /* constant expression evaluation / constant folding */
3166 if (get_opt_constant_folding()) {
3167 /* neither constants nor Tuple values can be evaluated */
3168 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
3169 /* try to evaluate */
3170 tv = computed_value(n);
3171 if (tv != tarval_bad) {
3173 type *old_tp = get_irn_type(n);
3174 int i, arity = get_irn_arity(n);
3178 * Try to recover the type of the new expression.
3180 for (i = 0; i < arity && !old_tp; ++i)
3181 old_tp = get_irn_type(get_irn_n(n, i));
3184 * we MUST copy the node here temporary, because it's still needed
3185 * for DBG_OPT_CSTEVAL
3187 node_size = offsetof(ir_node, attr) + n->op->attr_size;
3188 oldn = alloca(node_size);
3190 memcpy(oldn, n, node_size);
3191 CLONE_ARR_A(ir_node *, oldn->in, n->in);
3193 /* ARG, copy the in array, we need it for statistics */
3194 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
3196 /* note the inplace edges module */
3197 edges_node_deleted(n, current_ir_graph);
3199 /* evaluation was successful -- replace the node. */
3200 obstack_free(current_ir_graph->obst, n);
3201 nw = new_Const(get_tarval_mode (tv), tv);
3203 if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv))
3204 set_Const_type(nw, old_tp);
3205 DBG_OPT_CSTEVAL(oldn, nw);
3211 /* remove unnecessary nodes */
3212 if (get_opt_constant_folding() ||
3213 (iro == iro_Phi) || /* always optimize these nodes. */
3215 (iro == iro_Proj) ||
3216 (iro == iro_Block) ) /* Flags tested local. */
3217 n = equivalent_node (n);
3219 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
3221 /* Common Subexpression Elimination.
3223 * Checks whether n is already available.
3224 * The block input is used to distinguish different subexpressions. Right
3225 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
3226 * subexpressions within a block.
3229 n = identify_cons (current_ir_graph->value_table, n);
3232 edges_node_deleted(oldn, current_ir_graph);
3234 /* We found an existing, better node, so we can deallocate the old node. */
3235 obstack_free (current_ir_graph->obst, oldn);
3240 /* Some more constant expression evaluation that does not allow to
3242 iro = get_irn_opcode(n);
3243 if (get_opt_constant_folding() ||
3244 (iro == iro_Cond) ||
3245 (iro == iro_Proj) ||
3246 (iro == iro_Sel)) /* Flags tested local. */
3247 n = transform_node (n);
3249 /* Remove nodes with dead (Bad) input.
3250 Run always for transformation induced Bads. */
3253 /* Now we have a legal, useful node. Enter it in hash table for CSE */
3254 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
3255 n = identify_remember (current_ir_graph->value_table, n);
3263 * These optimizations never deallocate nodes (in place). This can cause dead
3264 * nodes lying on the obstack. Remove these by a dead node elimination,
3265 * i.e., a copying garbage collection.
3268 optimize_in_place_2 (ir_node *n)
3272 opcode iro = get_irn_opcode(n);
3274 if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n;
3276 /* constant expression evaluation / constant folding */
3277 if (get_opt_constant_folding()) {
3278 /* neither constants nor Tuple values can be evaluated */
3279 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
3280 /* try to evaluate */
3281 tv = computed_value(n);
3282 if (tv != tarval_bad) {
3283 /* evaluation was successful -- replace the node. */
3284 type *old_tp = get_irn_type(n);
3285 int i, arity = get_irn_arity(n);
3288 * Try to recover the type of the new expression.
3290 for (i = 0; i < arity && !old_tp; ++i)
3291 old_tp = get_irn_type(get_irn_n(n, i));
3293 n = new_Const(get_tarval_mode(tv), tv);
3295 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
3296 set_Const_type(n, old_tp);
3298 DBG_OPT_CSTEVAL(oldn, n);
3304 /* remove unnecessary nodes */
3305 if (get_opt_constant_folding() ||
3306 (iro == iro_Phi) || /* always optimize these nodes. */
3307 (iro == iro_Id) || /* ... */
3308 (iro == iro_Proj) || /* ... */
3309 (iro == iro_Block) ) /* Flags tested local. */
3310 n = equivalent_node(n);
3312 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
3314 /** common subexpression elimination **/
3315 /* Checks whether n is already available. */
3316 /* The block input is used to distinguish different subexpressions. Right
3317 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
3318 subexpressions within a block. */
3319 if (get_opt_cse()) {
3320 n = identify(current_ir_graph->value_table, n);
3323 /* Some more constant expression evaluation. */
3324 iro = get_irn_opcode(n);
3325 if (get_opt_constant_folding() ||
3326 (iro == iro_Cond) ||
3327 (iro == iro_Proj) ||
3328 (iro == iro_Sel)) /* Flags tested local. */
3329 n = transform_node(n);
3331 /* Remove nodes with dead (Bad) input.
3332 Run always for transformation induced Bads. */
3335 /* Now we can verify the node, as it has no dead inputs any more. */
3338 /* Now we have a legal, useful node. Enter it in hash table for cse.
3339 Blocks should be unique anyways. (Except the successor of start:
3340 is cse with the start block!) */
3341 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
3342 n = identify_remember(current_ir_graph->value_table, n);
3348 * Wrapper for external use, set proper status bits after optimization.
3351 optimize_in_place (ir_node *n)
3353 /* Handle graph state */
3354 assert(get_irg_phase_state(current_ir_graph) != phase_building);
3356 if (get_opt_global_cse())
3357 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3358 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
3359 set_irg_outs_inconsistent(current_ir_graph);
3361 /* Maybe we could also test whether optimizing the node can
3362 change the control graph. */
3363 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
3364 set_irg_dom_inconsistent(current_ir_graph);
3365 return optimize_in_place_2 (n);
3369 * set the default ir op operations
3371 ir_op *firm_set_default_operations(ir_op *op)
3373 op = firm_set_default_computed_value(op);
3374 op = firm_set_default_equivalent_node(op);
3375 op = firm_set_default_transform_node(op);
3376 op = firm_set_default_node_cmp_attr(op);
3377 op = firm_set_default_get_type(op);