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->ops.computed_value)
595 return n->op->ops.computed_value(n);
600 * set the default computed_value evaluator in an ir_op_ops.
602 * @param code the opcode for the default operation
603 * @param ops the operations initialized
608 static ir_op_ops *firm_set_default_computed_value(opcode code, ir_op_ops *ops)
612 ops->computed_value = computed_value_##a; \
647 /* returns 1 if the a and b are pointers to different locations. */
649 different_identity (ir_node *a, ir_node *b)
651 assert (mode_is_reference(get_irn_mode (a))
652 && mode_is_reference(get_irn_mode (b)));
654 if (get_irn_op (a) == op_Proj && get_irn_op(b) == op_Proj) {
655 ir_node *a1 = get_Proj_pred (a);
656 ir_node *b1 = get_Proj_pred (b);
657 if (a1 != b1 && get_irn_op (a1) == op_Alloc
658 && get_irn_op (b1) == op_Alloc)
666 * Returns a equivalent block for another block.
667 * If the block has only one predecessor, this is
668 * the equivalent one. If the only predecessor of a block is
669 * the block itself, this is a dead block.
671 * If both predecessors of a block are the branches of a binary
672 * Cond, the equivalent block is Cond's block.
674 * If all predecessors of a block are bad or lies in a dead
675 * block, the current block is dead as well.
677 * Note, that blocks are NEVER turned into Bad's, instead
678 * the dead_block flag is set. So, never test for is_Bad(block),
679 * always use is_dead_Block(block).
681 static ir_node *equivalent_node_Block(ir_node *n)
684 int n_preds = get_Block_n_cfgpreds(n);
686 /* The Block constructor does not call optimize, but mature_immBlock
687 calls the optimization. */
688 assert(get_Block_matured(n));
690 /* Straightening: a single entry Block following a single exit Block
691 can be merged, if it is not the Start block. */
692 /* !!! Beware, all Phi-nodes of n must have been optimized away.
693 This should be true, as the block is matured before optimize is called.
694 But what about Phi-cycles with the Phi0/Id that could not be resolved?
695 Remaining Phi nodes are just Ids. */
696 if ((n_preds == 1) && (get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp)) {
697 ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0));
698 if (predblock == oldn) {
699 /* Jmp jumps into the block it is in -- deal self cycle. */
700 n = set_Block_dead(n);
701 DBG_OPT_DEAD_BLOCK(oldn, n);
702 } else if (get_opt_control_flow_straightening()) {
704 DBG_OPT_STG(oldn, n);
707 else if ((n_preds == 1) &&
708 (get_irn_op(skip_Proj(get_Block_cfgpred(n, 0))) == op_Cond)) {
709 ir_node *predblock = get_Block_cfgpred_block(n, 0);
710 if (predblock == oldn) {
711 /* Jmp jumps into the block it is in -- deal self cycle. */
712 n = set_Block_dead(n);
713 DBG_OPT_DEAD_BLOCK(oldn, n);
716 else if ((n_preds == 2) &&
717 (get_opt_control_flow_weak_simplification())) {
718 /* Test whether Cond jumps twice to this block
719 * The more general case which more than 2 predecessors is handles
720 * in optimize_cf(), we handle only this special case for speed here.
722 ir_node *a = get_Block_cfgpred(n, 0);
723 ir_node *b = get_Block_cfgpred(n, 1);
725 if ((get_irn_op(a) == op_Proj) &&
726 (get_irn_op(b) == op_Proj) &&
727 (get_Proj_pred(a) == get_Proj_pred(b)) &&
728 (get_irn_op(get_Proj_pred(a)) == op_Cond) &&
729 (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) {
730 /* Also a single entry Block following a single exit Block. Phis have
731 twice the same operand and will be optimized away. */
732 n = get_nodes_block(get_Proj_pred(a));
733 DBG_OPT_IFSIM1(oldn, a, b, n);
736 else if (get_opt_unreachable_code() &&
737 (n != current_ir_graph->start_block) &&
738 (n != current_ir_graph->end_block) ) {
741 /* If all inputs are dead, this block is dead too, except if it is
742 the start or end block. This is one step of unreachable code
744 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
745 ir_node *pred = get_Block_cfgpred(n, i);
748 if (is_Bad(pred)) continue;
749 pred_blk = get_nodes_block(skip_Proj(pred));
751 if (is_Block_dead(pred_blk)) continue;
754 /* really found a living input */
759 n = set_Block_dead(n);
760 DBG_OPT_DEAD_BLOCK(oldn, n);
768 * Returns a equivalent node for a Jmp, a Bad :-)
769 * Of course this only happens if the Block of the Jmp is Bad.
771 static ir_node *equivalent_node_Jmp(ir_node *n)
773 /* unreachable code elimination */
774 if (is_Block_dead(get_nodes_block(n)))
780 /* Same for op_Raise */
781 #define equivalent_node_Raise equivalent_node_Jmp
784 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
785 See transform_node_Proj_Cond(). */
788 * optimize operations that are commutative and have neutral 0,
789 * so a op 0 = 0 op a = a.
791 static ir_node *equivalent_node_neutral_zero(ir_node *n)
795 ir_node *a = get_binop_left(n);
796 ir_node *b = get_binop_right(n);
801 /* After running compute_node there is only one constant predecessor.
802 Find this predecessors value and remember the other node: */
803 if ((tv = value_of(a)) != tarval_bad) {
805 } else if ((tv = value_of(b)) != tarval_bad) {
810 /* If this predecessors constant value is zero, the operation is
811 * unnecessary. Remove it.
813 * Beware: If n is a Add, the mode of on and n might be different
814 * which happens in this rare construction: NULL + 3.
815 * Then, a Conv would be needed which we cannot include here.
817 if (classify_tarval (tv) == TV_CLASSIFY_NULL) {
818 if (get_irn_mode(on) == get_irn_mode(n)) {
821 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
828 #define equivalent_node_Eor equivalent_node_neutral_zero
831 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
833 * The second one looks strange, but this construct
834 * is used heavily in the LCC sources :-).
836 * Beware: The Mode of an Add may be different than the mode of its
837 * predecessors, so we could not return a predecessors in all cases.
839 static ir_node *equivalent_node_Add(ir_node *n)
842 ir_node *left, *right;
844 n = equivalent_node_neutral_zero(n);
848 left = get_Add_left(n);
849 right = get_Add_right(n);
851 if (get_irn_op(left) == op_Sub) {
852 if (get_Sub_right(left) == right) {
855 n = get_Sub_left(left);
856 if (get_irn_mode(oldn) == get_irn_mode(n)) {
857 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
862 if (get_irn_op(right) == op_Sub) {
863 if (get_Sub_right(right) == left) {
866 n = get_Sub_left(right);
867 if (get_irn_mode(oldn) == get_irn_mode(n)) {
868 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
877 * optimize operations that are not commutative but have neutral 0 on left,
880 static ir_node *equivalent_node_left_zero(ir_node *n)
884 ir_node *a = get_binop_left(n);
885 ir_node *b = get_binop_right(n);
887 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
890 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
896 #define equivalent_node_Shl equivalent_node_left_zero
897 #define equivalent_node_Shr equivalent_node_left_zero
898 #define equivalent_node_Shrs equivalent_node_left_zero
899 #define equivalent_node_Rot equivalent_node_left_zero
902 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
904 * The second one looks strange, but this construct
905 * is used heavily in the LCC sources :-).
907 * Beware: The Mode of a Sub may be different than the mode of its
908 * predecessors, so we could not return a predecessors in all cases.
910 static ir_node *equivalent_node_Sub(ir_node *n)
914 ir_node *a = get_Sub_left(n);
915 ir_node *b = get_Sub_right(n);
917 /* Beware: modes might be different */
918 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
919 if (get_irn_mode(n) == get_irn_mode(a)) {
922 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
925 else if (get_irn_op(a) == op_Add) {
926 ir_mode *mode = get_irn_mode(n);
928 if (mode_wrap_around(mode)) {
929 ir_node *left = get_Add_left(a);
930 ir_node *right = get_Add_right(a);
933 if (get_irn_mode(n) == get_irn_mode(right)) {
935 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
938 else if (right == b) {
939 if (get_irn_mode(n) == get_irn_mode(left)) {
941 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
952 * Optimize an "idempotent unary op", ie op(op(n)) = n.
954 * @fixme -(-a) == a, but might overflow two times.
955 * We handle it anyway here but the better way would be a
956 * flag. This would be needed for Pascal for instance.
958 static ir_node *equivalent_node_idempotent_unop(ir_node *n)
961 ir_node *pred = get_unop_op(n);
963 /* optimize symmetric unop */
964 if (get_irn_op(pred) == get_irn_op(n)) {
965 n = get_unop_op(pred);
966 DBG_OPT_ALGSIM2(oldn, pred, n);
971 /* Not(Not(x)) == x */
972 #define equivalent_node_Not equivalent_node_idempotent_unop
974 /* --x == x */ /* ??? Is this possible or can --x raise an
975 out of bounds exception if min =! max? */
976 #define equivalent_node_Minus equivalent_node_idempotent_unop
979 * Optimize a * 1 = 1 * a = a.
981 static ir_node *equivalent_node_Mul(ir_node *n)
985 ir_node *a = get_Mul_left(n);
986 ir_node *b = get_Mul_right(n);
988 /* Mul is commutative and has again an other neutral element. */
989 if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) {
991 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
992 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) {
994 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1000 * Optimize a / 1 = a.
1002 static ir_node *equivalent_node_Div(ir_node *n)
1004 ir_node *a = get_Div_left(n);
1005 ir_node *b = get_Div_right(n);
1007 /* Div is not commutative. */
1008 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1009 /* Turn Div into a tuple (mem, bad, a) */
1010 ir_node *mem = get_Div_mem(n);
1011 turn_into_tuple(n, 3);
1012 set_Tuple_pred(n, pn_Div_M, mem);
1013 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
1014 set_Tuple_pred(n, pn_Div_res, a);
1020 * Optimize a / 1 = a.
1022 static ir_node *equivalent_node_DivMod(ir_node *n)
1024 ir_node *a = get_DivMod_left(n);
1025 ir_node *b = get_DivMod_right(n);
1027 /* Div is not commutative. */
1028 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1029 /* Turn DivMod into a tuple (mem, bad, a, 0) */
1030 ir_node *mem = get_Div_mem(n);
1031 ir_mode *mode = get_irn_mode(b);
1033 turn_into_tuple(n, 4);
1034 set_Tuple_pred(n, pn_DivMod_M, mem);
1035 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1036 set_Tuple_pred(n, pn_DivMod_res_div, a);
1037 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1043 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1045 static ir_node *equivalent_node_Or(ir_node *n)
1049 ir_node *a = get_Or_left(n);
1050 ir_node *b = get_Or_right(n);
1053 n = a; /* Or has it's own neutral element */
1054 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1055 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) {
1057 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1058 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
1060 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1067 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a.
1069 static ir_node *equivalent_node_And(ir_node *n)
1073 ir_node *a = get_And_left(n);
1074 ir_node *b = get_And_right(n);
1077 n = a; /* And has it's own neutral element */
1078 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1079 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) {
1081 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1082 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) {
1084 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1090 * Try to remove useless Conv's:
1092 static ir_node *equivalent_node_Conv(ir_node *n)
1095 ir_node *a = get_Conv_op(n);
1098 ir_mode *n_mode = get_irn_mode(n);
1099 ir_mode *a_mode = get_irn_mode(a);
1101 if (n_mode == a_mode) { /* No Conv necessary */
1103 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1104 } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */
1108 n_mode = get_irn_mode(n);
1109 b_mode = get_irn_mode(b);
1111 if (n_mode == b_mode) {
1112 if (n_mode == mode_b) {
1113 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1114 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1116 else if (mode_is_int(n_mode) || mode_is_character(n_mode)) {
1117 if (smaller_mode(b_mode, a_mode)){
1118 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1119 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1128 * A Cast may be removed if the type of the previous node
1129 * is already the type of the Cast.
1131 static ir_node *equivalent_node_Cast(ir_node *n) {
1133 ir_node *pred = get_Cast_op(n);
1135 if (get_irn_type(pred) == get_Cast_type(n)) {
1137 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1142 /* Several optimizations:
1143 - no Phi in start block.
1144 - remove Id operators that are inputs to Phi
1145 - fold Phi-nodes, iff they have only one predecessor except
1148 static ir_node *equivalent_node_Phi(ir_node *n)
1153 ir_node *block = NULL; /* to shutup gcc */
1154 ir_node *first_val = NULL; /* to shutup gcc */
1155 ir_node *scnd_val = NULL; /* to shutup gcc */
1157 if (!get_opt_normalize()) return n;
1159 n_preds = get_Phi_n_preds(n);
1161 block = get_nodes_block(n);
1162 /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
1163 assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
1164 if ((is_Block_dead(block)) || /* Control dead */
1165 (block == current_ir_graph->start_block)) /* There should be no Phi nodes */
1166 return new_Bad(); /* in the Start Block. */
1168 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1170 /* If the Block has a Bad pred, we also have one. */
1171 for (i = 0; i < n_preds; ++i)
1172 if (is_Bad(get_Block_cfgpred(block, i)))
1173 set_Phi_pred(n, i, new_Bad());
1175 /* Find first non-self-referencing input */
1176 for (i = 0; i < n_preds; ++i) {
1177 first_val = get_Phi_pred(n, i);
1178 if ( (first_val != n) /* not self pointer */
1180 && (! is_Bad(first_val))
1182 ) { /* value not dead */
1183 break; /* then found first value. */
1188 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1194 /* follow_Id () for rest of inputs, determine if any of these
1195 are non-self-referencing */
1196 while (++i < n_preds) {
1197 scnd_val = get_Phi_pred(n, i);
1198 if ( (scnd_val != n)
1199 && (scnd_val != first_val)
1201 && (! is_Bad(scnd_val))
1209 /* Fold, if no multiple distinct non-self-referencing inputs */
1211 DBG_OPT_PHI(oldn, n);
1213 /* skip the remaining Ids (done in get_Phi_pred). */
1214 /* superfluous, since we walk all to propagate Block's Bads.
1215 while (++i < n_preds) get_Phi_pred(n, i); */
1221 * optimize Proj(Tuple) and gigo() for ProjX in Bad block
1223 static ir_node *equivalent_node_Proj(ir_node *n)
1227 ir_node *a = get_Proj_pred(n);
1229 if ( get_irn_op(a) == op_Tuple) {
1230 /* Remove the Tuple/Proj combination. */
1231 if ( get_Proj_proj(n) <= get_Tuple_n_preds(a) ) {
1232 n = get_Tuple_pred(a, get_Proj_proj(n));
1233 DBG_OPT_TUPLE(oldn, a, n);
1235 assert(0); /* This should not happen! */
1239 else if (get_irn_mode(n) == mode_X) {
1240 if (is_Block_dead(get_nodes_block(skip_Proj(n)))) {
1241 /* Remove dead control flow -- early gigo(). */
1252 static ir_node *equivalent_node_Id(ir_node *n)
1258 } while (get_irn_op(n) == op_Id);
1260 DBG_OPT_ID(oldn, n);
1267 static ir_node *equivalent_node_Mux(ir_node *n)
1269 ir_node *oldn = n, *sel = get_Mux_sel(n);
1270 tarval *ts = value_of(sel);
1272 /* Mux(true, f, t) == t */
1273 if (ts == tarval_b_true) {
1274 n = get_Mux_true(n);
1275 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1277 /* Mux(false, f, t) == f */
1278 else if (ts == tarval_b_false) {
1279 n = get_Mux_false(n);
1280 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1282 /* Mux(v, x, x) == x */
1283 else if (get_Mux_false(n) == get_Mux_true(n)) {
1284 n = get_Mux_true(n);
1285 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1287 else if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(get_irn_mode(n))) {
1288 ir_node *cmp = get_Proj_pred(sel);
1289 long proj_nr = get_Proj_proj(sel);
1290 ir_node *b = get_Mux_false(n);
1291 ir_node *a = get_Mux_true(n);
1294 * Note: normalization puts the constant on the right site,
1295 * so we check only one case.
1297 * Note further that these optimization work even for floating point
1298 * with NaN's because -NaN == NaN.
1299 * However, if +0 and -0 is handled differently, we cannot use the first one.
1301 if (get_irn_op(cmp) == op_Cmp && get_Cmp_left(cmp) == a) {
1302 if (classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
1303 /* Mux(a CMP 0, X, a) */
1304 if (get_irn_op(b) == op_Minus && get_Minus_op(b) == a) {
1305 /* Mux(a CMP 0, -a, a) */
1306 if (proj_nr == pn_Cmp_Eq) {
1307 /* Mux(a == 0, -a, a) ==> -a */
1309 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1311 else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1312 /* Mux(a != 0, -a, a) ==> a */
1314 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1317 else if (classify_Const(b) == CNST_NULL) {
1318 /* Mux(a CMP 0, 0, a) */
1319 if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1320 /* Mux(a != 0, 0, a) ==> a */
1322 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1324 else if (proj_nr == pn_Cmp_Eq) {
1325 /* Mux(a == 0, 0, a) ==> 0 */
1327 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1337 * Optimize -a CMP -b into b CMP a.
1338 * This works only for for modes where unary Minus
1340 * Note that two-complement integers can Overflow
1341 * so it will NOT work.
1343 static ir_node *equivalent_node_Cmp(ir_node *n)
1345 ir_node *left = get_Cmp_left(n);
1346 ir_node *right = get_Cmp_right(n);
1348 if (get_irn_op(left) == op_Minus && get_irn_op(right) == op_Minus &&
1349 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1350 left = get_Minus_op(left);
1351 right = get_Minus_op(right);
1352 set_Cmp_left(n, right);
1353 set_Cmp_right(n, left);
1359 * Remove Confirm nodes if setting is on.
1360 * Replace Confirms(x, '=', Constlike) by Constlike.
1362 static ir_node *equivalent_node_Confirm(ir_node *n)
1364 ir_node *pred = get_Confirm_value(n);
1365 pn_Cmp pnc = get_Confirm_cmp(n);
1367 if (get_irn_op(pred) == op_Confirm && pnc == get_Confirm_cmp(pred)) {
1369 * rare case: two identical Confirms one after another,
1370 * replace the second one with the first.
1374 if (pnc == pn_Cmp_Eq) {
1375 ir_node *bound = get_Confirm_bound(n);
1378 * Optimize a rare case:
1379 * Confirm(x, '=', Constlike) ==> Constlike
1381 if (is_irn_constlike(bound)) {
1382 DBG_OPT_CONFIRM(n, bound);
1386 return get_opt_remove_Confirm() ? get_Confirm_value(n) : n;
1390 * Optimize CopyB(mem, x, x) into a Nop
1392 static ir_node *equivalent_node_CopyB(ir_node *n)
1394 ir_node *a = get_CopyB_dst(n);
1395 ir_node *b = get_CopyB_src(n);
1398 /* Turn CopyB into a tuple (mem, bad, bad) */
1399 ir_node *mem = get_CopyB_mem(n);
1400 turn_into_tuple(n, pn_CopyB_max);
1401 set_Tuple_pred(n, pn_CopyB_M, mem);
1402 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1403 set_Tuple_pred(n, pn_Call_M_except, new_Bad());
1409 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1410 * perform no actual computation, as, e.g., the Id nodes. It does not create
1411 * new nodes. It is therefore safe to free n if the node returned is not n.
1412 * If a node returns a Tuple we can not just skip it. If the size of the
1413 * in array fits, we transform n into a tuple (e.g., Div).
1416 equivalent_node(ir_node *n)
1418 if (n->op->ops.equivalent_node)
1419 return n->op->ops.equivalent_node(n);
1424 * sets the default equivalent node operation for an ir_op_ops.
1426 * @param code the opcode for the default operation
1427 * @param ops the operations initialized
1432 static ir_op_ops *firm_set_default_equivalent_node(opcode code, ir_op_ops *ops)
1436 ops->equivalent_node = equivalent_node_##a; \
1475 * Do node specific optimizations of nodes predecessors.
1478 optimize_preds(ir_node *n) {
1479 ir_node *a = NULL, *b = NULL;
1481 /* get the operands we will work on for simple cases. */
1483 a = get_binop_left(n);
1484 b = get_binop_right(n);
1485 } else if (is_unop(n)) {
1489 switch (get_irn_opcode(n)) {
1492 /* We don't want Cast as input to Cmp. */
1493 if (get_irn_op(a) == op_Cast) {
1497 if (get_irn_op(b) == op_Cast) {
1499 set_Cmp_right(n, b);
1508 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1509 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1510 * If possible, remove the Conv's.
1512 static ir_node *transform_node_AddSub(ir_node *n)
1514 ir_mode *mode = get_irn_mode(n);
1516 if (mode_is_reference(mode)) {
1517 ir_node *left = get_binop_left(n);
1518 ir_node *right = get_binop_right(n);
1519 int ref_bits = get_mode_size_bits(mode);
1521 if (get_irn_op(left) == op_Conv) {
1522 ir_mode *mode = get_irn_mode(left);
1523 int bits = get_mode_size_bits(mode);
1525 if (ref_bits == bits &&
1526 mode_is_int(mode) &&
1527 get_mode_arithmetic(mode) == irma_twos_complement) {
1528 ir_node *pre = get_Conv_op(left);
1529 ir_mode *pre_mode = get_irn_mode(pre);
1531 if (mode_is_int(pre_mode) &&
1532 get_mode_size_bits(pre_mode) == bits &&
1533 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1534 /* ok, this conv just changes to sign, moreover the calculation
1535 * is done with same number of bits as our address mode, so
1536 * we can ignore the conv as address calculation can be viewed
1537 * as either signed or unsigned
1539 set_binop_left(n, pre);
1544 if (get_irn_op(right) == op_Conv) {
1545 ir_mode *mode = get_irn_mode(right);
1546 int bits = get_mode_size_bits(mode);
1548 if (ref_bits == bits &&
1549 mode_is_int(mode) &&
1550 get_mode_arithmetic(mode) == irma_twos_complement) {
1551 ir_node *pre = get_Conv_op(right);
1552 ir_mode *pre_mode = get_irn_mode(pre);
1554 if (mode_is_int(pre_mode) &&
1555 get_mode_size_bits(pre_mode) == bits &&
1556 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1557 /* ok, this conv just changes to sign, moreover the calculation
1558 * is done with same number of bits as our address mode, so
1559 * we can ignore the conv as address calculation can be viewed
1560 * as either signed or unsigned
1562 set_binop_right(n, pre);
1571 * Do the AddSub optimization, then Transform Add(a,a) into Mul(a, 2)
1572 * if the mode is integer or float.
1573 * Transform Add(a,-b) into Sub(a,b).
1574 * Reassociation might fold this further.
1576 static ir_node *transform_node_Add(ir_node *n)
1581 n = transform_node_AddSub(n);
1583 mode = get_irn_mode(n);
1584 if (mode_is_num(mode)) {
1585 ir_node *a = get_Add_left(n);
1587 if (a == get_Add_right(n)) {
1588 ir_node *block = get_irn_n(n, -1);
1591 get_irn_dbg_info(n),
1595 new_r_Const_long(current_ir_graph, block, mode, 2),
1597 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
1600 ir_node *b = get_Add_right(n);
1602 if (get_irn_op(a) == op_Minus) {
1604 get_irn_dbg_info(n),
1610 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1612 else if (get_irn_op(b) == op_Minus) {
1614 get_irn_dbg_info(n),
1620 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1628 * Do the AddSub optimization, then Transform Sub(0,a) into Minus(a).
1630 static ir_node *transform_node_Sub(ir_node *n)
1635 n = transform_node_AddSub(n);
1637 mode = get_irn_mode(n);
1638 if (mode_is_num(mode) && (classify_Const(get_Sub_left(n)) == CNST_NULL)) {
1640 get_irn_dbg_info(n),
1645 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
1652 * Transform Mul(a,-1) into -a.
1653 * Do architecture dependent optimizations on Mul nodes
1655 static ir_node *transform_node_Mul(ir_node *n) {
1657 ir_mode *mode = get_irn_mode(n);
1659 if (mode_is_signed(mode)) {
1661 ir_node *a = get_Mul_left(n);
1662 ir_node *b = get_Mul_right(n);
1664 if (value_of(a) == get_mode_minus_one(mode))
1666 else if (value_of(b) == get_mode_minus_one(mode))
1669 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
1670 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
1674 return arch_dep_replace_mul_with_shifts(n);
1678 * transform a Div Node
1680 static ir_node *transform_node_Div(ir_node *n)
1682 tarval *tv = value_of(n);
1685 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
1687 if (tv != tarval_bad) {
1688 value = new_Const(get_tarval_mode(tv), tv);
1690 DBG_OPT_CSTEVAL(n, value);
1692 else /* Try architecture dependent optimization */
1693 value = arch_dep_replace_div_by_const(n);
1696 /* Turn Div into a tuple (mem, bad, value) */
1697 ir_node *mem = get_Div_mem(n);
1699 turn_into_tuple(n, pn_Div_max);
1700 set_Tuple_pred(n, pn_Div_M, mem);
1701 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
1702 set_Tuple_pred(n, pn_Div_res, value);
1708 * transform a Mod node
1710 static ir_node *transform_node_Mod(ir_node *n)
1712 tarval *tv = value_of(n);
1715 /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */
1717 if (tv != tarval_bad) {
1718 value = new_Const(get_tarval_mode(tv), tv);
1720 DBG_OPT_CSTEVAL(n, value);
1722 else /* Try architecture dependent optimization */
1723 value = arch_dep_replace_mod_by_const(n);
1726 /* Turn Mod into a tuple (mem, bad, value) */
1727 ir_node *mem = get_Mod_mem(n);
1729 turn_into_tuple(n, 3);
1730 set_Tuple_pred(n, pn_Mod_M, mem);
1731 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
1732 set_Tuple_pred(n, pn_Mod_res, value);
1738 * transform a DivMod node
1740 static ir_node *transform_node_DivMod(ir_node *n)
1744 ir_node *a = get_DivMod_left(n);
1745 ir_node *b = get_DivMod_right(n);
1746 ir_mode *mode = get_irn_mode(a);
1747 tarval *ta = value_of(a);
1748 tarval *tb = value_of(b);
1750 if (!(mode_is_int(mode) && mode_is_int(get_irn_mode(b))))
1753 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
1755 if (tb != tarval_bad) {
1756 if (tb == get_mode_one(get_tarval_mode(tb))) {
1757 b = new_Const (mode, get_mode_null(mode));
1760 DBG_OPT_CSTEVAL(n, b);
1762 else if (ta != tarval_bad) {
1763 tarval *resa, *resb;
1764 resa = tarval_div (ta, tb);
1765 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
1766 Jmp for X result!? */
1767 resb = tarval_mod (ta, tb);
1768 if (resb == tarval_bad) return n; /* Causes exception! */
1769 a = new_Const (mode, resa);
1770 b = new_Const (mode, resb);
1773 DBG_OPT_CSTEVAL(n, a);
1774 DBG_OPT_CSTEVAL(n, b);
1776 else { /* Try architecture dependent optimization */
1777 arch_dep_replace_divmod_by_const(&a, &b, n);
1778 evaluated = a != NULL;
1780 } else if (ta == get_mode_null(mode)) {
1781 /* 0 / non-Const = 0 */
1786 if (evaluated) { /* replace by tuple */
1787 ir_node *mem = get_DivMod_mem(n);
1788 turn_into_tuple(n, 4);
1789 set_Tuple_pred(n, pn_DivMod_M, mem);
1790 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1791 set_Tuple_pred(n, pn_DivMod_res_div, a);
1792 set_Tuple_pred(n, pn_DivMod_res_mod, b);
1799 * Optimize Abs(x) into x if x is Confirmed >= 0
1800 * Optimize Abs(x) into -x if x is Confirmed <= 0
1802 static ir_node *transform_node_Abs(ir_node *n)
1805 ir_node *a = get_Abs_op(n);
1806 value_classify sign = classify_value_sign(a);
1808 if (sign == VALUE_NEGATIVE) {
1809 ir_mode *mode = get_irn_mode(n);
1812 * We can replace the Abs by -x here.
1813 * We even could add a new Confirm here.
1815 * Note that -x would create a new node, so we could
1816 * not run it in the equivalent_node() context.
1818 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
1819 get_irn_n(n, -1), a, mode);
1821 DBG_OPT_CONFIRM(oldn, n);
1823 else if (sign == VALUE_POSITIVE) {
1824 /* n is positive, Abs is not needed */
1827 DBG_OPT_CONFIRM(oldn, n);
1834 * transform a Cond node
1836 static ir_node *transform_node_Cond(ir_node *n)
1838 /* Replace the Cond by a Jmp if it branches on a constant
1841 ir_node *a = get_Cond_selector(n);
1842 tarval *ta = value_of(a);
1844 /* we need block info which is not available in floating irgs */
1845 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
1848 if ((ta != tarval_bad) &&
1849 (get_irn_mode(a) == mode_b) &&
1850 (get_opt_unreachable_code())) {
1851 /* It's a boolean Cond, branching on a boolean constant.
1852 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
1853 jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n));
1854 turn_into_tuple(n, 2);
1855 if (ta == tarval_b_true) {
1856 set_Tuple_pred(n, pn_Cond_false, new_Bad());
1857 set_Tuple_pred(n, pn_Cond_true, jmp);
1859 set_Tuple_pred(n, pn_Cond_false, jmp);
1860 set_Tuple_pred(n, pn_Cond_true, new_Bad());
1862 /* We might generate an endless loop, so keep it alive. */
1863 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
1871 static ir_node *transform_node_Eor(ir_node *n)
1874 ir_node *a = get_Eor_left(n);
1875 ir_node *b = get_Eor_right(n);
1876 ir_mode *mode = get_irn_mode(n);
1880 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
1881 mode, get_mode_null(mode));
1882 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
1884 else if ((mode == mode_b)
1885 && (get_irn_op(a) == op_Proj)
1886 && (get_irn_mode(a) == mode_b)
1887 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)
1888 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
1889 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
1890 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
1891 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
1893 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
1895 else if ((mode == mode_b)
1896 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) {
1897 /* The Eor is a Not. Replace it by a Not. */
1898 /* ????!!!Extend to bitfield 1111111. */
1899 n = new_r_Not(current_ir_graph, get_irn_n(n, -1), a, mode_b);
1901 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
1908 * Transform a boolean Not.
1910 static ir_node *transform_node_Not(ir_node *n)
1913 ir_node *a = get_Not_op(n);
1915 if ( (get_irn_mode(n) == mode_b)
1916 && (get_irn_op(a) == op_Proj)
1917 && (get_irn_mode(a) == mode_b)
1918 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
1919 /* We negate a Cmp. The Cmp has the negated result anyways! */
1920 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
1921 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
1922 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
1929 * Transform a Cast_type(Const) into a new Const_type
1931 static ir_node *transform_node_Cast(ir_node *n) {
1933 ir_node *pred = get_Cast_op(n);
1934 type *tp = get_irn_type(n);
1936 if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) {
1937 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
1938 get_Const_tarval(pred), tp);
1939 DBG_OPT_CSTEVAL(oldn, n);
1940 } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) {
1941 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred),
1942 get_SymConst_kind(pred), tp);
1943 DBG_OPT_CSTEVAL(oldn, n);
1950 * Transform a Proj(Div) with a non-zero value.
1951 * Removes the exceptions and routes the memory to the NoMem node.
1953 static ir_node *transform_node_Proj_Div(ir_node *proj)
1955 ir_node *n = get_Proj_pred(proj);
1956 ir_node *b = get_Div_right(n);
1959 if (value_not_zero(b)) {
1960 /* div(x, y) && y != 0 */
1961 proj_nr = get_Proj_proj(proj);
1963 /* this node may float */
1964 set_irn_pinned(n, op_pin_state_floats);
1966 if (proj_nr == pn_Div_X_except) {
1967 /* we found an exception handler, remove it */
1969 } else if (proj_nr == pn_Div_M) {
1970 /* the memory Proj can be removed */
1971 ir_node *res = get_Div_mem(n);
1972 set_Div_mem(n, get_irg_no_mem(current_ir_graph));
1981 * Transform a Proj(Mod) with a non-zero value.
1982 * Removes the exceptions and routes the memory to the NoMem node.
1984 static ir_node *transform_node_Proj_Mod(ir_node *proj)
1986 ir_node *n = get_Proj_pred(proj);
1987 ir_node *b = get_Mod_right(n);
1990 if (value_not_zero(b)) {
1991 /* mod(x, y) && y != 0 */
1992 proj_nr = get_Proj_proj(proj);
1994 /* this node may float */
1995 set_irn_pinned(n, op_pin_state_floats);
1997 if (proj_nr == pn_Mod_X_except) {
1998 /* we found an exception handler, remove it */
2000 } else if (proj_nr == pn_Mod_M) {
2001 /* the memory Proj can be removed */
2002 ir_node *res = get_Mod_mem(n);
2003 set_Mod_mem(n, get_irg_no_mem(current_ir_graph));
2007 else if (proj_nr == pn_Mod_res && get_Mod_left(n) == b) {
2008 /* a % a = 0 if a != 0 */
2009 ir_mode *mode = get_irn_mode(proj);
2010 ir_node *res = new_Const(mode, get_mode_null(mode));
2012 DBG_OPT_CSTEVAL(n, res);
2020 * Transform a Proj(DivMod) with a non-zero value.
2021 * Removes the exceptions and routes the memory to the NoMem node.
2023 static ir_node *transform_node_Proj_DivMod(ir_node *proj)
2025 ir_node *n = get_Proj_pred(proj);
2026 ir_node *b = get_DivMod_right(n);
2029 if (value_not_zero(b)) {
2030 /* DivMod(x, y) && y != 0 */
2031 proj_nr = get_Proj_proj(proj);
2033 /* this node may float */
2034 set_irn_pinned(n, op_pin_state_floats);
2036 if (proj_nr == pn_DivMod_X_except) {
2037 /* we found an exception handler, remove it */
2040 else if (proj_nr == pn_DivMod_M) {
2041 /* the memory Proj can be removed */
2042 ir_node *res = get_DivMod_mem(n);
2043 set_DivMod_mem(n, get_irg_no_mem(current_ir_graph));
2047 else if (proj_nr == pn_DivMod_res_mod && get_DivMod_left(n) == b) {
2048 /* a % a = 0 if a != 0 */
2049 ir_mode *mode = get_irn_mode(proj);
2050 ir_node *res = new_Const(mode, get_mode_null(mode));
2052 DBG_OPT_CSTEVAL(n, res);
2060 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
2062 static ir_node *transform_node_Proj_Cond(ir_node *proj)
2064 if (get_opt_unreachable_code()) {
2065 ir_node *n = get_Proj_pred(proj);
2066 ir_node *b = get_Cond_selector(n);
2068 if (mode_is_int(get_irn_mode(b))) {
2069 tarval *tb = value_of(b);
2071 if (tb != tarval_bad) {
2072 /* we have a constant switch */
2073 long num = get_Proj_proj(proj);
2075 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
2076 if (get_tarval_long(tb) == num) {
2077 /* Do NOT create a jump here, or we will have 2 control flow ops
2078 * in a block. This case is optimized away in optimize_cf(). */
2082 /* this case will NEVER be taken, kill it */
2093 * Normalizes and optimizes Cmp nodes.
2095 static ir_node *transform_node_Proj_Cmp(ir_node *proj)
2097 if (get_opt_reassociation()) {
2098 ir_node *n = get_Proj_pred(proj);
2099 ir_node *left = get_Cmp_left(n);
2100 ir_node *right = get_Cmp_right(n);
2104 ir_mode *mode = NULL;
2105 long proj_nr = get_Proj_proj(proj);
2108 * First step: normalize the compare op
2109 * by placing the constant on the right site
2110 * or moving the lower address node to the left.
2111 * We ignore the case that both are constants
2112 * this case should be optimized away.
2114 if (get_irn_op(right) == op_Const)
2116 else if (get_irn_op(left) == op_Const) {
2121 proj_nr = get_inversed_pnc(proj_nr);
2124 else if (left > right) {
2130 proj_nr = get_inversed_pnc(proj_nr);
2135 * Second step: Try to reduce the magnitude
2136 * of a constant. This may help to generate better code
2137 * later and may help to normalize more compares.
2138 * Of course this is only possible for integer values.
2141 mode = get_irn_mode(c);
2142 tv = get_Const_tarval(c);
2144 if (tv != tarval_bad) {
2145 /* the following optimization is possible on modes without Overflow
2146 * on Unary Minus or on == and !=:
2147 * -a CMP c ==> a swap(CMP) -c
2149 * Beware: for two-complement Overflow may occur, so only == and != can
2150 * be optimized, see this:
2151 * -MININT < 0 =/=> MININT > 0 !!!
2153 if (get_opt_constant_folding() && get_irn_op(left) == op_Minus &&
2154 (!mode_overflow_on_unary_Minus(mode) ||
2155 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
2156 left = get_Minus_op(left);
2157 tv = tarval_sub(get_mode_null(mode), tv);
2159 proj_nr = get_inversed_pnc(proj_nr);
2163 /* for integer modes, we have more */
2164 if (mode_is_int(mode)) {
2165 /* Ne includes Unordered which is not possible on integers.
2166 * However, frontends often use this wrong, so fix it here */
2167 if (proj_nr & pn_Cmp_Uo) {
2168 proj_nr &= ~pn_Cmp_Uo;
2169 set_Proj_proj(proj, proj_nr);
2172 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
2173 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
2174 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
2175 tv = tarval_sub(tv, get_mode_one(mode));
2177 proj_nr ^= pn_Cmp_Eq;
2180 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
2181 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
2182 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
2183 tv = tarval_add(tv, get_mode_one(mode));
2185 proj_nr ^= pn_Cmp_Eq;
2189 /* the following reassociations work only for == and != */
2190 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
2192 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
2193 if (classify_tarval(tv) == TV_CLASSIFY_NULL && get_irn_op(left) == op_Sub) {
2194 right = get_Sub_right(left);
2195 left = get_Sub_left(left);
2197 tv = value_of(right);
2201 if (tv != tarval_bad) {
2202 ir_op *op = get_irn_op(left);
2204 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
2206 ir_node *c1 = get_Sub_right(left);
2207 tarval *tv2 = value_of(c1);
2209 if (tv2 != tarval_bad) {
2210 tv2 = tarval_add(tv, value_of(c1));
2212 if (tv2 != tarval_bad) {
2213 left = get_Sub_left(left);
2219 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
2220 else if (op == op_Add) {
2221 ir_node *a_l = get_Add_left(left);
2222 ir_node *a_r = get_Add_right(left);
2226 if (get_irn_op(a_l) == op_Const) {
2228 tv2 = value_of(a_l);
2232 tv2 = value_of(a_r);
2235 if (tv2 != tarval_bad) {
2236 tv2 = tarval_sub(tv, tv2);
2238 if (tv2 != tarval_bad) {
2245 /* -a == c ==> a == -c, -a != c ==> a != -c */
2246 else if (op == op_Minus) {
2247 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
2249 if (tv2 != tarval_bad) {
2250 left = get_Minus_op(left);
2257 /* the following reassociations work only for <= */
2258 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2259 if (tv != tarval_bad) {
2260 ir_op *op = get_irn_op(left);
2262 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
2272 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
2274 if (changed & 2) /* need a new Const */
2275 right = new_Const(mode, tv);
2277 /* create a new compare */
2278 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block,
2281 set_Proj_pred(proj, n);
2282 set_Proj_proj(proj, proj_nr);
2289 * Does all optimizations on nodes that must be done on it's Proj's
2290 * because of creating new nodes.
2292 static ir_node *transform_node_Proj(ir_node *proj)
2294 ir_node *n = get_Proj_pred(proj);
2296 switch (get_irn_opcode(n)) {
2298 return transform_node_Proj_Div(proj);
2301 return transform_node_Proj_Mod(proj);
2304 return transform_node_Proj_DivMod(proj);
2307 return transform_node_Proj_Cond(proj);
2310 return transform_node_Proj_Cmp(proj);
2313 /* should not happen, but if it does will be optimized away */
2314 return equivalent_node_Proj(proj);
2323 * returns the operands of a commutative bin-op, if one operand is
2324 * a const, it is returned as the second one.
2326 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c)
2328 ir_node *op_a = get_binop_left(binop);
2329 ir_node *op_b = get_binop_right(binop);
2331 assert(is_op_commutative(get_irn_op(binop)));
2333 if (get_irn_op(op_a) == op_Const) {
2344 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
2345 * Such pattern may arise in bitfield stores.
2347 * value c4 value c4 & c2
2348 * AND c3 AND c1 | c3
2353 static ir_node *transform_node_Or_bf_store(ir_node *or)
2357 ir_node *and_l, *c3;
2358 ir_node *value, *c4;
2359 ir_node *new_and, *new_const, *block;
2360 ir_mode *mode = get_irn_mode(or);
2362 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
2364 get_comm_Binop_Ops(or, &and, &c1);
2365 if ((get_irn_op(c1) != op_Const) || (get_irn_op(and) != op_And))
2368 get_comm_Binop_Ops(and, &or_l, &c2);
2369 if ((get_irn_op(c2) != op_Const) || (get_irn_op(or_l) != op_Or))
2372 get_comm_Binop_Ops(or_l, &and_l, &c3);
2373 if ((get_irn_op(c3) != op_Const) || (get_irn_op(and_l) != op_And))
2376 get_comm_Binop_Ops(and_l, &value, &c4);
2377 if (get_irn_op(c4) != op_Const)
2380 /* ok, found the pattern, check for conditions */
2381 assert(mode == get_irn_mode(and));
2382 assert(mode == get_irn_mode(or_l));
2383 assert(mode == get_irn_mode(and_l));
2385 tv1 = get_Const_tarval(c1);
2386 tv2 = get_Const_tarval(c2);
2387 tv3 = get_Const_tarval(c3);
2388 tv4 = get_Const_tarval(c4);
2390 tv = tarval_or(tv4, tv2);
2391 if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) {
2392 /* have at least one 0 at the same bit position */
2396 n_tv4 = tarval_not(tv4);
2397 if (tv3 != tarval_and(tv3, n_tv4)) {
2398 /* bit in the or_mask is outside the and_mask */
2402 n_tv2 = tarval_not(tv2);
2403 if (tv1 != tarval_and(tv1, n_tv2)) {
2404 /* bit in the or_mask is outside the and_mask */
2408 /* ok, all conditions met */
2409 block = get_irn_n(or, -1);
2411 new_and = new_r_And(current_ir_graph, block,
2412 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
2414 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
2416 set_Or_left(or, new_and);
2417 set_Or_right(or, new_const);
2419 /* check for more */
2420 return transform_node_Or_bf_store(or);
2424 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
2426 static ir_node *transform_node_Or_Rot(ir_node *or)
2428 ir_mode *mode = get_irn_mode(or);
2429 ir_node *shl, *shr, *block;
2430 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
2433 if (! mode_is_int(mode))
2436 shl = get_binop_left(or);
2437 shr = get_binop_right(or);
2439 if (get_irn_op(shl) == op_Shr) {
2440 if (get_irn_op(shr) != op_Shl)
2447 else if (get_irn_op(shl) != op_Shl)
2449 else if (get_irn_op(shr) != op_Shr)
2452 x = get_Shl_left(shl);
2453 if (x != get_Shr_left(shr))
2456 c1 = get_Shl_right(shl);
2457 c2 = get_Shr_right(shr);
2458 if (get_irn_op(c1) == op_Const && get_irn_op(c2) == op_Const) {
2459 tv1 = get_Const_tarval(c1);
2460 if (! tarval_is_long(tv1))
2463 tv2 = get_Const_tarval(c2);
2464 if (! tarval_is_long(tv2))
2467 if (get_tarval_long(tv1) + get_tarval_long(tv2)
2468 != get_mode_size_bits(mode))
2471 /* yet, condition met */
2472 block = get_irn_n(or, -1);
2474 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
2476 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
2479 else if (get_irn_op(c1) == op_Sub) {
2483 if (get_Sub_right(sub) != v)
2486 c1 = get_Sub_left(sub);
2487 if (get_irn_op(c1) != op_Const)
2490 tv1 = get_Const_tarval(c1);
2491 if (! tarval_is_long(tv1))
2494 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
2497 /* yet, condition met */
2498 block = get_nodes_block(or);
2500 /* a Rot right is not supported, so use a rot left */
2501 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
2503 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
2506 else if (get_irn_op(c2) == op_Sub) {
2510 c1 = get_Sub_left(sub);
2511 if (get_irn_op(c1) != op_Const)
2514 tv1 = get_Const_tarval(c1);
2515 if (! tarval_is_long(tv1))
2518 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
2521 /* yet, condition met */
2522 block = get_irn_n(or, -1);
2525 n = new_r_Rot(current_ir_graph, block, x, v, mode);
2527 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
2537 static ir_node *transform_node_Or(ir_node *or)
2539 or = transform_node_Or_bf_store(or);
2540 or = transform_node_Or_Rot(or);
2546 static ir_node *transform_node(ir_node *n);
2549 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl.
2551 * Should be moved to reassociation?
2553 static ir_node *transform_node_shift(ir_node *n)
2555 ir_node *left, *right;
2556 tarval *tv1, *tv2, *res;
2558 int modulo_shf, flag;
2560 left = get_binop_left(n);
2562 /* different operations */
2563 if (get_irn_op(left) != get_irn_op(n))
2566 right = get_binop_right(n);
2567 tv1 = value_of(right);
2568 if (tv1 == tarval_bad)
2571 tv2 = value_of(get_binop_right(left));
2572 if (tv2 == tarval_bad)
2575 res = tarval_add(tv1, tv2);
2577 /* beware: a simple replacement works only, if res < modulo shift */
2578 mode = get_irn_mode(n);
2582 modulo_shf = get_mode_modulo_shift(mode);
2583 if (modulo_shf > 0) {
2584 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
2586 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
2593 /* ok, we can replace it */
2594 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
2596 in[0] = get_binop_left(left);
2597 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
2599 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
2601 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
2603 return transform_node(irn);
2608 #define transform_node_Shr transform_node_shift
2609 #define transform_node_Shrs transform_node_shift
2610 #define transform_node_Shl transform_node_shift
2613 * Remove dead blocks and nodes in dead blocks
2614 * in keep alive list. We do not generate a new End node.
2616 static ir_node *transform_node_End(ir_node *n) {
2617 int i, n_keepalives = get_End_n_keepalives(n);
2619 for (i = 0; i < n_keepalives; ++i) {
2620 ir_node *ka = get_End_keepalive(n, i);
2622 if (is_Block_dead(ka)) {
2623 set_End_keepalive(n, i, new_Bad());
2626 else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka)))
2627 set_End_keepalive(n, i, new_Bad());
2633 * Optimize a Mux into some simpler cases.
2635 static ir_node *transform_node_Mux(ir_node *n)
2637 ir_node *oldn = n, *sel = get_Mux_sel(n);
2638 ir_mode *mode = get_irn_mode(n);
2640 if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(mode)) {
2641 ir_node *cmp = get_Proj_pred(sel);
2642 long proj_nr = get_Proj_proj(sel);
2643 ir_node *f = get_Mux_false(n);
2644 ir_node *t = get_Mux_true(n);
2646 if (get_irn_op(cmp) == op_Cmp && classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
2647 ir_node *block = get_irn_n(n, -1);
2650 * Note: normalization puts the constant on the right site,
2651 * so we check only one case.
2653 * Note further that these optimization work even for floating point
2654 * with NaN's because -NaN == NaN.
2655 * However, if +0 and -0 is handled differently, we cannot use the first one.
2657 if (get_irn_op(f) == op_Minus &&
2658 get_Minus_op(f) == t &&
2659 get_Cmp_left(cmp) == t) {
2661 if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
2662 /* Mux(a >=/> 0, -a, a) ==> Abs(a) */
2663 n = new_rd_Abs(get_irn_dbg_info(n),
2667 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2670 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2671 /* Mux(a <=/< 0, -a, a) ==> Minus(Abs(a)) */
2672 n = new_rd_Abs(get_irn_dbg_info(n),
2676 n = new_rd_Minus(get_irn_dbg_info(n),
2681 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2685 else if (get_irn_op(t) == op_Minus &&
2686 get_Minus_op(t) == f &&
2687 get_Cmp_left(cmp) == f) {
2689 if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2690 /* Mux(a <=/< 0, a, -a) ==> Abs(a) */
2691 n = new_rd_Abs(get_irn_dbg_info(n),
2695 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2698 else if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
2699 /* Mux(a >=/> 0, a, -a) ==> Minus(Abs(a)) */
2700 n = new_rd_Abs(get_irn_dbg_info(n),
2704 n = new_rd_Minus(get_irn_dbg_info(n),
2709 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2714 if (mode_is_int(mode) && mode_is_signed(mode) &&
2715 get_mode_arithmetic(mode) == irma_twos_complement) {
2716 ir_node *x = get_Cmp_left(cmp);
2718 /* the following optimization works only with signed integer two-complement mode */
2720 if (mode == get_irn_mode(x)) {
2722 * FIXME: this restriction is two rigid, as it would still
2723 * work if mode(x) = Hs and mode == Is, but at least it removes
2726 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Le) &&
2727 classify_Const(t) == CNST_ALL_ONE &&
2728 classify_Const(f) == CNST_NULL) {
2730 * Mux(x:T </<= 0, 0, -1) -> Shrs(x, sizeof_bits(T) - 1)
2734 n = new_rd_Shrs(get_irn_dbg_info(n),
2735 current_ir_graph, block, x,
2736 new_r_Const_long(current_ir_graph, block, mode_Iu,
2737 get_mode_size_bits(mode) - 1),
2739 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
2742 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) &&
2743 classify_Const(t) == CNST_ONE &&
2744 classify_Const(f) == CNST_NULL) {
2746 * Mux(x:T >/>= 0, 0, 1) -> Shr(-x, sizeof_bits(T) - 1)
2750 n = new_rd_Shr(get_irn_dbg_info(n),
2751 current_ir_graph, block,
2752 new_r_Minus(current_ir_graph, block, x, mode),
2753 new_r_Const_long(current_ir_graph, block, mode_Iu,
2754 get_mode_size_bits(mode) - 1),
2756 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
2763 return arch_transform_node_Mux(n);
2767 * Tries several [inplace] [optimizing] transformations and returns an
2768 * equivalent node. The difference to equivalent_node() is that these
2769 * transformations _do_ generate new nodes, and thus the old node must
2770 * not be freed even if the equivalent node isn't the old one.
2772 static ir_node *transform_node(ir_node *n)
2774 if (n->op->ops.transform_node)
2775 n = n->op->ops.transform_node(n);
2780 * sSets the default transform node operation for an ir_op_ops.
2782 * @param code the opcode for the default operation
2783 * @param ops the operations initialized
2788 static ir_op_ops *firm_set_default_transform_node(opcode code, ir_op_ops *ops)
2792 ops->transform_node = transform_node_##a; \
2824 /* **************** Common Subexpression Elimination **************** */
2826 /** The size of the hash table used, should estimate the number of nodes
2828 #define N_IR_NODES 512
2830 /** Compares the attributes of two Const nodes. */
2831 static int node_cmp_attr_Const(ir_node *a, ir_node *b)
2833 return (get_Const_tarval(a) != get_Const_tarval(b))
2834 || (get_Const_type(a) != get_Const_type(b));
2837 /** Compares the attributes of two Proj nodes. */
2838 static int node_cmp_attr_Proj(ir_node *a, ir_node *b)
2840 return get_irn_proj_attr (a) != get_irn_proj_attr (b);
2843 /** Compares the attributes of two Filter nodes. */
2844 static int node_cmp_attr_Filter(ir_node *a, ir_node *b)
2846 return get_Filter_proj(a) != get_Filter_proj(b);
2849 /** Compares the attributes of two Alloc nodes. */
2850 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b)
2852 return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where)
2853 || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type);
2856 /** Compares the attributes of two Free nodes. */
2857 static int node_cmp_attr_Free(ir_node *a, ir_node *b)
2859 return (get_irn_free_attr(a).where != get_irn_free_attr(b).where)
2860 || (get_irn_free_attr(a).type != get_irn_free_attr(b).type);
2863 /** Compares the attributes of two SymConst nodes. */
2864 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b)
2866 return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num)
2867 || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p)
2868 || (get_irn_symconst_attr(a).tp != get_irn_symconst_attr(b).tp);
2871 /** Compares the attributes of two Call nodes. */
2872 static int node_cmp_attr_Call(ir_node *a, ir_node *b)
2874 return (get_irn_call_attr(a) != get_irn_call_attr(b));
2877 /** Compares the attributes of two Sel nodes. */
2878 static int node_cmp_attr_Sel(ir_node *a, ir_node *b)
2880 return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind)
2881 || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name)
2882 || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner)
2883 || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name)
2884 || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type);
2887 /** Compares the attributes of two Phi nodes. */
2888 static int node_cmp_attr_Phi(ir_node *a, ir_node *b)
2890 return get_irn_phi_attr (a) != get_irn_phi_attr (b);
2893 /** Compares the attributes of two Cast nodes. */
2894 static int node_cmp_attr_Cast(ir_node *a, ir_node *b)
2896 return get_Cast_type(a) != get_Cast_type(b);
2899 /** Compares the attributes of two Load nodes. */
2900 static int node_cmp_attr_Load(ir_node *a, ir_node *b)
2902 if (get_Load_volatility(a) == volatility_is_volatile ||
2903 get_Load_volatility(b) == volatility_is_volatile)
2904 /* NEVER do CSE on volatile Loads */
2907 return get_Load_mode(a) != get_Load_mode(b);
2910 /** Compares the attributes of two Store nodes. */
2911 static int node_cmp_attr_Store(ir_node *a, ir_node *b)
2913 /* NEVER do CSE on volatile Stores */
2914 return (get_Store_volatility(a) == volatility_is_volatile ||
2915 get_Store_volatility(b) == volatility_is_volatile);
2918 /** Compares the attributes of two Confirm nodes. */
2919 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b)
2921 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
2925 * Set the default node attribute compare operation for an ir_op_ops.
2927 * @param code the opcode for the default operation
2928 * @param ops the operations initialized
2933 static ir_op_ops *firm_set_default_node_cmp_attr(opcode code, ir_op_ops *ops)
2937 ops->node_cmp_attr = node_cmp_attr_##a; \
2963 * Compare function for two nodes in the hash table. Gets two
2964 * nodes as parameters. Returns 0 if the nodes are a cse.
2967 vt_cmp (const void *elt, const void *key)
2975 if (a == b) return 0;
2977 if ((get_irn_op(a) != get_irn_op(b)) ||
2978 (get_irn_mode(a) != get_irn_mode(b))) return 1;
2980 /* compare if a's in and b's in are of equal length */
2981 irn_arity_a = get_irn_intra_arity (a);
2982 if (irn_arity_a != get_irn_intra_arity(b))
2985 /* for block-local cse and op_pin_state_pinned nodes: */
2986 if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) {
2987 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
2991 /* compare a->in[0..ins] with b->in[0..ins] */
2992 for (i = 0; i < irn_arity_a; i++)
2993 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
2997 * here, we already now that the nodes are identical except their
3000 if (a->op->ops.node_cmp_attr)
3001 return a->op->ops.node_cmp_attr(a, b);
3007 * Calculate a hash value of a node.
3010 ir_node_hash (ir_node *node)
3015 if (node->op == op_Const) {
3016 /* special value for const, as they only differ in their tarval. */
3017 h = HASH_PTR(node->attr.con.tv);
3018 h = 9*h + HASH_PTR(get_irn_mode(node));
3019 } else if (node->op == op_SymConst) {
3020 /* special value for const, as they only differ in their symbol. */
3021 h = HASH_PTR(node->attr.i.sym.type_p);
3022 h = 9*h + HASH_PTR(get_irn_mode(node));
3025 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
3026 h = irn_arity = get_irn_intra_arity(node);
3028 /* consider all in nodes... except the block if not a control flow. */
3029 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
3030 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
3034 h = 9*h + HASH_PTR(get_irn_mode(node));
3036 h = 9*h + HASH_PTR(get_irn_op(node));
3043 new_identities(void) {
3044 return new_pset(vt_cmp, N_IR_NODES);
3048 del_identities(pset *value_table) {
3049 del_pset(value_table);
3053 * Return the canonical node computing the same value as n.
3054 * Looks up the node in a hash table.
3056 * For Const nodes this is performed in the constructor, too. Const
3057 * nodes are extremely time critical because of their frequent use in
3058 * constant string arrays.
3060 static INLINE ir_node *
3061 identify (pset *value_table, ir_node *n)
3065 if (!value_table) return n;
3067 if (get_opt_reassociation()) {
3068 if (is_op_commutative(get_irn_op(n))) {
3069 ir_node *l = get_binop_left(n);
3070 ir_node *r = get_binop_right(n);
3072 /* for commutative operators perform a OP b == b OP a */
3074 set_binop_left(n, r);
3075 set_binop_right(n, l);
3080 o = pset_find (value_table, n, ir_node_hash (n));
3089 * During construction we set the op_pin_state_pinned flag in the graph right when the
3090 * optimization is performed. The flag turning on procedure global cse could
3091 * be changed between two allocations. This way we are safe.
3093 static INLINE ir_node *
3094 identify_cons (pset *value_table, ir_node *n) {
3097 n = identify(value_table, n);
3098 if (get_irn_n(old, -1) != get_irn_n(n, -1))
3099 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3104 * Return the canonical node computing the same value as n.
3105 * Looks up the node in a hash table, enters it in the table
3106 * if it isn't there yet.
3109 identify_remember (pset *value_table, ir_node *n)
3113 if (!value_table) return n;
3115 if (get_opt_reassociation()) {
3116 if (is_op_commutative(get_irn_op(n))) {
3117 ir_node *l = get_binop_left(n);
3118 ir_node *r = get_binop_right(n);
3120 /* for commutative operators perform a OP b == b OP a */
3122 set_binop_left(n, r);
3123 set_binop_right(n, l);
3128 /* lookup or insert in hash table with given hash key. */
3129 o = pset_insert (value_table, n, ir_node_hash (n));
3139 add_identities (pset *value_table, ir_node *node) {
3140 if (get_opt_cse() && (get_irn_opcode(node) != iro_Block))
3141 identify_remember (value_table, node);
3145 * garbage in, garbage out. If a node has a dead input, i.e., the
3146 * Bad node is input to the node, return the Bad node.
3148 static INLINE ir_node *
3149 gigo (ir_node *node)
3152 ir_op *op = get_irn_op(node);
3154 /* remove garbage blocks by looking at control flow that leaves the block
3155 and replacing the control flow by Bad. */
3156 if (get_irn_mode(node) == mode_X) {
3157 ir_node *block = get_nodes_block(skip_Proj(node));
3159 /* Don't optimize nodes in immature blocks. */
3160 if (!get_Block_matured(block)) return node;
3161 /* Don't optimize End, may have Bads. */
3162 if (op == op_End) return node;
3164 if (is_Block(block)) {
3165 irn_arity = get_irn_arity(block);
3166 for (i = 0; i < irn_arity; i++) {
3167 if (!is_Bad(get_irn_n(block, i)))
3170 if (i == irn_arity) return new_Bad();
3174 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
3175 blocks predecessors is dead. */
3176 if ( op != op_Block && op != op_Phi && op != op_Tuple) {
3177 irn_arity = get_irn_arity(node);
3180 * Beware: we can only read the block of a non-floating node.
3182 if (is_irn_pinned_in_irg(node) &&
3183 is_Block_dead(get_nodes_block(node)))
3186 for (i = 0; i < irn_arity; i++) {
3187 ir_node *pred = get_irn_n(node, i);
3191 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
3192 return new_Unknown(get_irn_mode(node));
3196 /* With this code we violate the agreement that local_optimize
3197 only leaves Bads in Block, Phi and Tuple nodes. */
3198 /* If Block has only Bads as predecessors it's garbage. */
3199 /* If Phi has only Bads as predecessors it's garbage. */
3200 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
3201 irn_arity = get_irn_arity(node);
3202 for (i = 0; i < irn_arity; i++) {
3203 if (!is_Bad(get_irn_n(node, i))) break;
3205 if (i == irn_arity) node = new_Bad();
3213 * These optimizations deallocate nodes from the obstack.
3214 * It can only be called if it is guaranteed that no other nodes
3215 * reference this one, i.e., right after construction of a node.
3217 * current_ir_graph must be set to the graph of the node!
3220 optimize_node(ir_node *n)
3224 opcode iro = get_irn_opcode(n);
3226 /* Always optimize Phi nodes: part of the construction. */
3227 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
3229 /* constant expression evaluation / constant folding */
3230 if (get_opt_constant_folding()) {
3231 /* neither constants nor Tuple values can be evaluated */
3232 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
3233 /* try to evaluate */
3234 tv = computed_value(n);
3235 if (tv != tarval_bad) {
3237 type *old_tp = get_irn_type(n);
3238 int i, arity = get_irn_arity(n);
3242 * Try to recover the type of the new expression.
3244 for (i = 0; i < arity && !old_tp; ++i)
3245 old_tp = get_irn_type(get_irn_n(n, i));
3248 * we MUST copy the node here temporary, because it's still needed
3249 * for DBG_OPT_CSTEVAL
3251 node_size = offsetof(ir_node, attr) + n->op->attr_size;
3252 oldn = alloca(node_size);
3254 memcpy(oldn, n, node_size);
3255 CLONE_ARR_A(ir_node *, oldn->in, n->in);
3257 /* ARG, copy the in array, we need it for statistics */
3258 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
3260 /* note the inplace edges module */
3261 edges_node_deleted(n, current_ir_graph);
3263 /* evaluation was successful -- replace the node. */
3264 obstack_free(current_ir_graph->obst, n);
3265 nw = new_Const(get_tarval_mode (tv), tv);
3267 if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv))
3268 set_Const_type(nw, old_tp);
3269 DBG_OPT_CSTEVAL(oldn, nw);
3275 /* remove unnecessary nodes */
3276 if (get_opt_constant_folding() ||
3277 (iro == iro_Phi) || /* always optimize these nodes. */
3279 (iro == iro_Proj) ||
3280 (iro == iro_Block) ) /* Flags tested local. */
3281 n = equivalent_node (n);
3283 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
3285 /* Common Subexpression Elimination.
3287 * Checks whether n is already available.
3288 * The block input is used to distinguish different subexpressions. Right
3289 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
3290 * subexpressions within a block.
3293 n = identify_cons (current_ir_graph->value_table, n);
3296 edges_node_deleted(oldn, current_ir_graph);
3298 /* We found an existing, better node, so we can deallocate the old node. */
3299 obstack_free (current_ir_graph->obst, oldn);
3304 /* Some more constant expression evaluation that does not allow to
3306 iro = get_irn_opcode(n);
3307 if (get_opt_constant_folding() ||
3308 (iro == iro_Cond) ||
3309 (iro == iro_Proj) ||
3310 (iro == iro_Sel)) /* Flags tested local. */
3311 n = transform_node (n);
3313 /* Remove nodes with dead (Bad) input.
3314 Run always for transformation induced Bads. */
3317 /* Now we have a legal, useful node. Enter it in hash table for CSE */
3318 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
3319 n = identify_remember (current_ir_graph->value_table, n);
3327 * These optimizations never deallocate nodes (in place). This can cause dead
3328 * nodes lying on the obstack. Remove these by a dead node elimination,
3329 * i.e., a copying garbage collection.
3332 optimize_in_place_2 (ir_node *n)
3336 opcode iro = get_irn_opcode(n);
3338 if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n;
3340 /* constant expression evaluation / constant folding */
3341 if (get_opt_constant_folding()) {
3342 /* neither constants nor Tuple values can be evaluated */
3343 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
3344 /* try to evaluate */
3345 tv = computed_value(n);
3346 if (tv != tarval_bad) {
3347 /* evaluation was successful -- replace the node. */
3348 type *old_tp = get_irn_type(n);
3349 int i, arity = get_irn_arity(n);
3352 * Try to recover the type of the new expression.
3354 for (i = 0; i < arity && !old_tp; ++i)
3355 old_tp = get_irn_type(get_irn_n(n, i));
3357 n = new_Const(get_tarval_mode(tv), tv);
3359 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
3360 set_Const_type(n, old_tp);
3362 DBG_OPT_CSTEVAL(oldn, n);
3368 /* remove unnecessary nodes */
3369 if (get_opt_constant_folding() ||
3370 (iro == iro_Phi) || /* always optimize these nodes. */
3371 (iro == iro_Id) || /* ... */
3372 (iro == iro_Proj) || /* ... */
3373 (iro == iro_Block) ) /* Flags tested local. */
3374 n = equivalent_node(n);
3376 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
3378 /** common subexpression elimination **/
3379 /* Checks whether n is already available. */
3380 /* The block input is used to distinguish different subexpressions. Right
3381 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
3382 subexpressions within a block. */
3383 if (get_opt_cse()) {
3384 n = identify(current_ir_graph->value_table, n);
3387 /* Some more constant expression evaluation. */
3388 iro = get_irn_opcode(n);
3389 if (get_opt_constant_folding() ||
3390 (iro == iro_Cond) ||
3391 (iro == iro_Proj) ||
3392 (iro == iro_Sel)) /* Flags tested local. */
3393 n = transform_node(n);
3395 /* Remove nodes with dead (Bad) input.
3396 Run always for transformation induced Bads. */
3399 /* Now we can verify the node, as it has no dead inputs any more. */
3402 /* Now we have a legal, useful node. Enter it in hash table for cse.
3403 Blocks should be unique anyways. (Except the successor of start:
3404 is cse with the start block!) */
3405 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
3406 n = identify_remember(current_ir_graph->value_table, n);
3412 * Wrapper for external use, set proper status bits after optimization.
3415 optimize_in_place (ir_node *n)
3417 /* Handle graph state */
3418 assert(get_irg_phase_state(current_ir_graph) != phase_building);
3420 if (get_opt_global_cse())
3421 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3422 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
3423 set_irg_outs_inconsistent(current_ir_graph);
3425 /* Maybe we could also test whether optimizing the node can
3426 change the control graph. */
3427 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
3428 set_irg_dom_inconsistent(current_ir_graph);
3429 return optimize_in_place_2 (n);
3433 * Sets the default operation for an ir_ops.
3435 ir_op_ops *firm_set_default_operations(opcode code, ir_op_ops *ops)
3437 ops = firm_set_default_computed_value(code, ops);
3438 ops = firm_set_default_equivalent_node(code, ops);
3439 ops = firm_set_default_transform_node(code, ops);
3440 ops = firm_set_default_node_cmp_attr(code, ops);
3441 ops = firm_set_default_get_type(code, ops);