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(get_Proj_pred(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) ) {
735 /* If all inputs are dead, this block is dead too, except if it is
736 the start or end block. This is one step of unreachable code
738 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --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);
754 DBG_OPT_DEAD_BLOCK(oldn, n);
762 * Returns a equivalent node for a Jmp, a Bad :-)
763 * Of course this only happens if the Block of the Jmp is Bad.
765 static ir_node *equivalent_node_Jmp(ir_node *n)
767 /* unreachable code elimination */
768 if (is_Block_dead(get_nodes_block(n)))
774 /* Same for op_Raise */
775 #define equivalent_node_Raise equivalent_node_Jmp
778 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
779 See transform_node_Proj_Cond(). */
782 * optimize operations that are commutative and have neutral 0,
783 * so a op 0 = 0 op a = a.
785 static ir_node *equivalent_node_neutral_zero(ir_node *n)
789 ir_node *a = get_binop_left(n);
790 ir_node *b = get_binop_right(n);
795 /* After running compute_node there is only one constant predecessor.
796 Find this predecessors value and remember the other node: */
797 if ((tv = value_of(a)) != tarval_bad) {
799 } else if ((tv = value_of(b)) != tarval_bad) {
804 /* If this predecessors constant value is zero, the operation is
805 * unnecessary. Remove it.
807 * Beware: If n is a Add, the mode of on and n might be different
808 * which happens in this rare construction: NULL + 3.
809 * Then, a Conv would be needed which we cannot include here.
811 if (classify_tarval (tv) == TV_CLASSIFY_NULL) {
812 if (get_irn_mode(on) == get_irn_mode(n)) {
815 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
822 #define equivalent_node_Eor equivalent_node_neutral_zero
825 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
827 * The second one looks strange, but this construct
828 * is used heavily in the LCC sources :-).
830 * Beware: The Mode of an Add may be different than the mode of its
831 * predecessors, so we could not return a predecessors in all cases.
833 static ir_node *equivalent_node_Add(ir_node *n)
836 ir_node *left, *right;
838 n = equivalent_node_neutral_zero(n);
842 left = get_Add_left(n);
843 right = get_Add_right(n);
845 if (get_irn_op(left) == op_Sub) {
846 if (get_Sub_right(left) == right) {
849 n = get_Sub_left(left);
850 if (get_irn_mode(oldn) == get_irn_mode(n)) {
851 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
856 if (get_irn_op(right) == op_Sub) {
857 if (get_Sub_right(right) == left) {
860 n = get_Sub_left(right);
861 if (get_irn_mode(oldn) == get_irn_mode(n)) {
862 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
871 * optimize operations that are not commutative but have neutral 0 on left,
874 static ir_node *equivalent_node_left_zero(ir_node *n)
878 ir_node *a = get_binop_left(n);
879 ir_node *b = get_binop_right(n);
881 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
884 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
890 #define equivalent_node_Shl equivalent_node_left_zero
891 #define equivalent_node_Shr equivalent_node_left_zero
892 #define equivalent_node_Shrs equivalent_node_left_zero
893 #define equivalent_node_Rot equivalent_node_left_zero
896 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
898 * The second one looks strange, but this construct
899 * is used heavily in the LCC sources :-).
901 * Beware: The Mode of a Sub may be different than the mode of its
902 * predecessors, so we could not return a predecessors in all cases.
904 static ir_node *equivalent_node_Sub(ir_node *n)
908 ir_node *a = get_Sub_left(n);
909 ir_node *b = get_Sub_right(n);
911 /* Beware: modes might be different */
912 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
913 if (get_irn_mode(n) == get_irn_mode(a)) {
916 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
919 else if (get_irn_op(a) == op_Add) {
920 ir_mode *mode = get_irn_mode(n);
922 if (mode_wrap_around(mode)) {
923 ir_node *left = get_Add_left(a);
924 ir_node *right = get_Add_right(a);
927 if (get_irn_mode(n) == get_irn_mode(right)) {
929 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
932 else if (right == b) {
933 if (get_irn_mode(n) == get_irn_mode(left)) {
935 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
946 * Optimize an "idempotent unary op", ie op(op(n)) = n.
948 * @fixme -(-a) == a, but might overflow two times.
949 * We handle it anyway here but the better way would be a
950 * flag. This would be needed for Pascal for instance.
952 static ir_node *equivalent_node_idempotent_unop(ir_node *n)
955 ir_node *pred = get_unop_op(n);
957 /* optimize symmetric unop */
958 if (get_irn_op(pred) == get_irn_op(n)) {
959 n = get_unop_op(pred);
960 DBG_OPT_ALGSIM2(oldn, pred, n);
965 /* Not(Not(x)) == x */
966 #define equivalent_node_Not equivalent_node_idempotent_unop
968 /* --x == x */ /* ??? Is this possible or can --x raise an
969 out of bounds exception if min =! max? */
970 #define equivalent_node_Minus equivalent_node_idempotent_unop
973 * Optimize a * 1 = 1 * a = a.
975 static ir_node *equivalent_node_Mul(ir_node *n)
979 ir_node *a = get_Mul_left(n);
980 ir_node *b = get_Mul_right(n);
982 /* Mul is commutative and has again an other neutral element. */
983 if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) {
985 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
986 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) {
988 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
994 * Optimize a / 1 = a.
996 static ir_node *equivalent_node_Div(ir_node *n)
998 ir_node *a = get_Div_left(n);
999 ir_node *b = get_Div_right(n);
1001 /* Div is not commutative. */
1002 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1003 /* Turn Div into a tuple (mem, bad, a) */
1004 ir_node *mem = get_Div_mem(n);
1005 turn_into_tuple(n, 3);
1006 set_Tuple_pred(n, pn_Div_M, mem);
1007 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
1008 set_Tuple_pred(n, pn_Div_res, a);
1014 * Optimize a / 1 = a.
1016 static ir_node *equivalent_node_DivMod(ir_node *n)
1018 ir_node *a = get_DivMod_left(n);
1019 ir_node *b = get_DivMod_right(n);
1021 /* Div is not commutative. */
1022 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1023 /* Turn DivMod into a tuple (mem, bad, a, 0) */
1024 ir_node *mem = get_Div_mem(n);
1025 ir_mode *mode = get_irn_mode(b);
1027 turn_into_tuple(n, 4);
1028 set_Tuple_pred(n, pn_DivMod_M, mem);
1029 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1030 set_Tuple_pred(n, pn_DivMod_res_div, a);
1031 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1037 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1039 static ir_node *equivalent_node_Or(ir_node *n)
1043 ir_node *a = get_Or_left(n);
1044 ir_node *b = get_Or_right(n);
1047 n = a; /* Or has it's own neutral element */
1048 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1049 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) {
1051 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1052 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
1054 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1061 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a.
1063 static ir_node *equivalent_node_And(ir_node *n)
1067 ir_node *a = get_And_left(n);
1068 ir_node *b = get_And_right(n);
1071 n = a; /* And has it's own neutral element */
1072 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1073 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) {
1075 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1076 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) {
1078 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1084 * Try to remove useless Conv's:
1086 static ir_node *equivalent_node_Conv(ir_node *n)
1089 ir_node *a = get_Conv_op(n);
1092 ir_mode *n_mode = get_irn_mode(n);
1093 ir_mode *a_mode = get_irn_mode(a);
1095 if (n_mode == a_mode) { /* No Conv necessary */
1097 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1098 } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */
1102 n_mode = get_irn_mode(n);
1103 b_mode = get_irn_mode(b);
1105 if (n_mode == b_mode) {
1106 if (n_mode == mode_b) {
1107 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1108 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1110 else if (mode_is_int(n_mode) || mode_is_character(n_mode)) {
1111 if (smaller_mode(b_mode, a_mode)){
1112 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1113 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1122 * A Cast may be removed if the type of the previous node
1123 * is already the type of the Cast.
1125 static ir_node *equivalent_node_Cast(ir_node *n) {
1127 ir_node *pred = get_Cast_op(n);
1129 if (get_irn_type(pred) == get_Cast_type(n)) {
1131 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1136 /* Several optimizations:
1137 - no Phi in start block.
1138 - remove Id operators that are inputs to Phi
1139 - fold Phi-nodes, iff they have only one predecessor except
1142 static ir_node *equivalent_node_Phi(ir_node *n)
1147 ir_node *block = NULL; /* to shutup gcc */
1148 ir_node *first_val = NULL; /* to shutup gcc */
1149 ir_node *scnd_val = NULL; /* to shutup gcc */
1151 if (!get_opt_normalize()) return n;
1153 n_preds = get_Phi_n_preds(n);
1155 block = get_nodes_block(n);
1156 /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
1157 assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
1158 if ((is_Block_dead(block)) || /* Control dead */
1159 (block == current_ir_graph->start_block)) /* There should be no Phi nodes */
1160 return new_Bad(); /* in the Start Block. */
1162 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1164 /* If the Block has a Bad pred, we also have one. */
1165 for (i = 0; i < n_preds; ++i)
1166 if (is_Bad(get_Block_cfgpred(block, i)))
1167 set_Phi_pred(n, i, new_Bad());
1169 /* Find first non-self-referencing input */
1170 for (i = 0; i < n_preds; ++i) {
1171 first_val = get_Phi_pred(n, i);
1172 if ( (first_val != n) /* not self pointer */
1174 && (! is_Bad(first_val))
1176 ) { /* value not dead */
1177 break; /* then found first value. */
1182 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1188 /* follow_Id () for rest of inputs, determine if any of these
1189 are non-self-referencing */
1190 while (++i < n_preds) {
1191 scnd_val = get_Phi_pred(n, i);
1192 if ( (scnd_val != n)
1193 && (scnd_val != first_val)
1195 && (! is_Bad(scnd_val))
1203 /* Fold, if no multiple distinct non-self-referencing inputs */
1205 DBG_OPT_PHI(oldn, n);
1207 /* skip the remaining Ids (done in get_Phi_pred). */
1208 /* superfluous, since we walk all to propagate Block's Bads.
1209 while (++i < n_preds) get_Phi_pred(n, i); */
1215 * optimize Proj(Tuple) and gigo() for ProjX in Bad block
1217 static ir_node *equivalent_node_Proj(ir_node *n)
1221 ir_node *a = get_Proj_pred(n);
1223 if ( get_irn_op(a) == op_Tuple) {
1224 /* Remove the Tuple/Proj combination. */
1225 if ( get_Proj_proj(n) <= get_Tuple_n_preds(a) ) {
1226 n = get_Tuple_pred(a, get_Proj_proj(n));
1227 DBG_OPT_TUPLE(oldn, a, n);
1229 assert(0); /* This should not happen! */
1233 else if (get_irn_mode(n) == mode_X) {
1234 if (is_Block_dead(get_nodes_block(skip_Proj(n)))) {
1235 /* Remove dead control flow -- early gigo(). */
1246 static ir_node *equivalent_node_Id(ir_node *n)
1252 } while (get_irn_op(n) == op_Id);
1254 DBG_OPT_ID(oldn, n);
1261 static ir_node *equivalent_node_Mux(ir_node *n)
1263 ir_node *oldn = n, *sel = get_Mux_sel(n);
1264 tarval *ts = value_of(sel);
1266 /* Mux(true, f, t) == t */
1267 if (ts == tarval_b_true) {
1268 n = get_Mux_true(n);
1269 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1271 /* Mux(false, f, t) == f */
1272 else if (ts == tarval_b_false) {
1273 n = get_Mux_false(n);
1274 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1276 /* Mux(v, x, x) == x */
1277 else if (get_Mux_false(n) == get_Mux_true(n)) {
1278 n = get_Mux_true(n);
1279 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1281 else if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(get_irn_mode(n))) {
1282 ir_node *cmp = get_Proj_pred(sel);
1283 long proj_nr = get_Proj_proj(sel);
1284 ir_node *b = get_Mux_false(n);
1285 ir_node *a = get_Mux_true(n);
1288 * Note: normalization puts the constant on the right site,
1289 * so we check only one case.
1291 * Note further that these optimization work even for floating point
1292 * with NaN's because -NaN == NaN.
1293 * However, if +0 and -0 is handled differently, we cannot use the first one.
1295 if (get_irn_op(cmp) == op_Cmp && get_Cmp_left(cmp) == a) {
1296 if (classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
1297 /* Mux(a CMP 0, X, a) */
1298 if (get_irn_op(b) == op_Minus && get_Minus_op(b) == a) {
1299 /* Mux(a CMP 0, -a, a) */
1300 if (proj_nr == pn_Cmp_Eq) {
1301 /* Mux(a == 0, -a, a) ==> -a */
1303 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1305 else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1306 /* Mux(a != 0, -a, a) ==> a */
1308 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1311 else if (classify_Const(b) == CNST_NULL) {
1312 /* Mux(a CMP 0, 0, a) */
1313 if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1314 /* Mux(a != 0, 0, a) ==> a */
1316 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1318 else if (proj_nr == pn_Cmp_Eq) {
1319 /* Mux(a == 0, 0, a) ==> 0 */
1321 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1331 * Optimize -a CMP -b into b CMP a.
1332 * This works only for for modes where unary Minus
1334 * Note that two-complement integers can Overflow
1335 * so it will NOT work.
1337 static ir_node *equivalent_node_Cmp(ir_node *n)
1339 ir_node *left = get_Cmp_left(n);
1340 ir_node *right = get_Cmp_right(n);
1342 if (get_irn_op(left) == op_Minus && get_irn_op(right) == op_Minus &&
1343 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1344 left = get_Minus_op(left);
1345 right = get_Minus_op(right);
1346 set_Cmp_left(n, right);
1347 set_Cmp_right(n, left);
1353 * Remove Confirm nodes if setting is on.
1354 * Replace Confirms(x, '=', Constlike) by Constlike.
1356 static ir_node *equivalent_node_Confirm(ir_node *n)
1358 ir_node *pred = get_Confirm_value(n);
1359 pn_Cmp pnc = get_Confirm_cmp(n);
1361 if (get_irn_op(pred) == op_Confirm && pnc == get_Confirm_cmp(pred)) {
1363 * rare case: two identical Confirms one after another,
1364 * replace the second one with the first.
1368 if (pnc == pn_Cmp_Eq) {
1369 ir_node *bound = get_Confirm_bound(n);
1372 * Optimize a rare case:
1373 * Confirm(x, '=', Constlike) ==> Constlike
1375 if (is_irn_constlike(bound)) {
1376 DBG_OPT_CONFIRM(n, bound);
1380 return get_opt_remove_Confirm() ? get_Confirm_value(n) : n;
1384 * Optimize CopyB(mem, x, x) into a Nop
1386 static ir_node *equivalent_node_CopyB(ir_node *n)
1388 ir_node *a = get_CopyB_dst(n);
1389 ir_node *b = get_CopyB_src(n);
1392 /* Turn CopyB into a tuple (mem, bad, bad) */
1393 ir_node *mem = get_CopyB_mem(n);
1394 turn_into_tuple(n, pn_CopyB_max);
1395 set_Tuple_pred(n, pn_CopyB_M, mem);
1396 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1397 set_Tuple_pred(n, pn_Call_M_except, new_Bad());
1403 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1404 * perform no actual computation, as, e.g., the Id nodes. It does not create
1405 * new nodes. It is therefore safe to free n if the node returned is not n.
1406 * If a node returns a Tuple we can not just skip it. If the size of the
1407 * in array fits, we transform n into a tuple (e.g., Div).
1410 equivalent_node(ir_node *n)
1412 if (n->op->equivalent_node)
1413 return n->op->equivalent_node(n);
1418 * set the default equivalent node operation
1420 static ir_op *firm_set_default_equivalent_node(ir_op *op)
1424 op->equivalent_node = equivalent_node_##a; \
1455 op->equivalent_node = NULL;
1463 * Do node specific optimizations of nodes predecessors.
1466 optimize_preds(ir_node *n) {
1467 ir_node *a = NULL, *b = NULL;
1469 /* get the operands we will work on for simple cases. */
1471 a = get_binop_left(n);
1472 b = get_binop_right(n);
1473 } else if (is_unop(n)) {
1477 switch (get_irn_opcode(n)) {
1480 /* We don't want Cast as input to Cmp. */
1481 if (get_irn_op(a) == op_Cast) {
1485 if (get_irn_op(b) == op_Cast) {
1487 set_Cmp_right(n, b);
1496 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1497 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1498 * If possible, remove the Conv's.
1500 static ir_node *transform_node_AddSub(ir_node *n)
1502 ir_mode *mode = get_irn_mode(n);
1504 if (mode_is_reference(mode)) {
1505 ir_node *left = get_binop_left(n);
1506 ir_node *right = get_binop_right(n);
1507 int ref_bits = get_mode_size_bits(mode);
1509 if (get_irn_op(left) == op_Conv) {
1510 ir_mode *mode = get_irn_mode(left);
1511 int bits = get_mode_size_bits(mode);
1513 if (ref_bits == bits &&
1514 mode_is_int(mode) &&
1515 get_mode_arithmetic(mode) == irma_twos_complement) {
1516 ir_node *pre = get_Conv_op(left);
1517 ir_mode *pre_mode = get_irn_mode(pre);
1519 if (mode_is_int(pre_mode) &&
1520 get_mode_size_bits(pre_mode) == bits &&
1521 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1522 /* ok, this conv just changes to sign, moreover the calculation
1523 * is done with same number of bits as our address mode, so
1524 * we can ignore the conv as address calculation can be viewed
1525 * as either signed or unsigned
1527 set_binop_left(n, pre);
1532 if (get_irn_op(right) == op_Conv) {
1533 ir_mode *mode = get_irn_mode(right);
1534 int bits = get_mode_size_bits(mode);
1536 if (ref_bits == bits &&
1537 mode_is_int(mode) &&
1538 get_mode_arithmetic(mode) == irma_twos_complement) {
1539 ir_node *pre = get_Conv_op(right);
1540 ir_mode *pre_mode = get_irn_mode(pre);
1542 if (mode_is_int(pre_mode) &&
1543 get_mode_size_bits(pre_mode) == bits &&
1544 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1545 /* ok, this conv just changes to sign, moreover the calculation
1546 * is done with same number of bits as our address mode, so
1547 * we can ignore the conv as address calculation can be viewed
1548 * as either signed or unsigned
1550 set_binop_right(n, pre);
1559 * Do the AddSub optimization, then Transform Add(a,a) into Mul(a, 2)
1560 * if the mode is integer or float.
1561 * Transform Add(a,-b) into Sub(a,b).
1562 * Reassociation might fold this further.
1564 static ir_node *transform_node_Add(ir_node *n)
1569 n = transform_node_AddSub(n);
1571 mode = get_irn_mode(n);
1572 if (mode_is_num(mode)) {
1573 ir_node *a = get_Add_left(n);
1575 if (a == get_Add_right(n)) {
1576 ir_node *block = get_irn_n(n, -1);
1579 get_irn_dbg_info(n),
1583 new_r_Const_long(current_ir_graph, block, mode, 2),
1585 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
1588 ir_node *b = get_Add_right(n);
1590 if (get_irn_op(a) == op_Minus) {
1592 get_irn_dbg_info(n),
1598 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1600 else if (get_irn_op(b) == op_Minus) {
1602 get_irn_dbg_info(n),
1608 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1616 * Do the AddSub optimization, then Transform Sub(0,a) into Minus(a).
1618 static ir_node *transform_node_Sub(ir_node *n)
1623 n = transform_node_AddSub(n);
1625 mode = get_irn_mode(n);
1626 if (mode_is_num(mode) && (classify_Const(get_Sub_left(n)) == CNST_NULL)) {
1628 get_irn_dbg_info(n),
1633 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
1640 * Transform Mul(a,-1) into -a.
1641 * Do architecture dependent optimizations on Mul nodes
1643 static ir_node *transform_node_Mul(ir_node *n) {
1645 ir_mode *mode = get_irn_mode(n);
1647 if (mode_is_signed(mode)) {
1649 ir_node *a = get_Mul_left(n);
1650 ir_node *b = get_Mul_right(n);
1652 if (value_of(a) == get_mode_minus_one(mode))
1654 else if (value_of(b) == get_mode_minus_one(mode))
1657 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
1658 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
1662 return arch_dep_replace_mul_with_shifts(n);
1666 * transform a Div Node
1668 static ir_node *transform_node_Div(ir_node *n)
1670 tarval *tv = value_of(n);
1673 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
1675 if (tv != tarval_bad) {
1676 value = new_Const(get_tarval_mode(tv), tv);
1678 DBG_OPT_CSTEVAL(n, value);
1680 else /* Try architecture dependent optimization */
1681 value = arch_dep_replace_div_by_const(n);
1684 /* Turn Div into a tuple (mem, bad, value) */
1685 ir_node *mem = get_Div_mem(n);
1687 turn_into_tuple(n, pn_Div_max);
1688 set_Tuple_pred(n, pn_Div_M, mem);
1689 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
1690 set_Tuple_pred(n, pn_Div_res, value);
1696 * transform a Mod node
1698 static ir_node *transform_node_Mod(ir_node *n)
1700 tarval *tv = value_of(n);
1703 /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */
1705 if (tv != tarval_bad) {
1706 value = new_Const(get_tarval_mode(tv), tv);
1708 DBG_OPT_CSTEVAL(n, value);
1710 else /* Try architecture dependent optimization */
1711 value = arch_dep_replace_mod_by_const(n);
1714 /* Turn Mod into a tuple (mem, bad, value) */
1715 ir_node *mem = get_Mod_mem(n);
1717 turn_into_tuple(n, 3);
1718 set_Tuple_pred(n, pn_Mod_M, mem);
1719 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
1720 set_Tuple_pred(n, pn_Mod_res, value);
1726 * transform a DivMod node
1728 static ir_node *transform_node_DivMod(ir_node *n)
1732 ir_node *a = get_DivMod_left(n);
1733 ir_node *b = get_DivMod_right(n);
1734 ir_mode *mode = get_irn_mode(a);
1735 tarval *ta = value_of(a);
1736 tarval *tb = value_of(b);
1738 if (!(mode_is_int(mode) && mode_is_int(get_irn_mode(b))))
1741 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
1743 if (tb != tarval_bad) {
1744 if (tb == get_mode_one(get_tarval_mode(tb))) {
1745 b = new_Const (mode, get_mode_null(mode));
1748 DBG_OPT_CSTEVAL(n, b);
1750 else if (ta != tarval_bad) {
1751 tarval *resa, *resb;
1752 resa = tarval_div (ta, tb);
1753 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
1754 Jmp for X result!? */
1755 resb = tarval_mod (ta, tb);
1756 if (resb == tarval_bad) return n; /* Causes exception! */
1757 a = new_Const (mode, resa);
1758 b = new_Const (mode, resb);
1761 DBG_OPT_CSTEVAL(n, a);
1762 DBG_OPT_CSTEVAL(n, b);
1764 else { /* Try architecture dependent optimization */
1765 arch_dep_replace_divmod_by_const(&a, &b, n);
1766 evaluated = a != NULL;
1768 } else if (ta == get_mode_null(mode)) {
1769 /* 0 / non-Const = 0 */
1774 if (evaluated) { /* replace by tuple */
1775 ir_node *mem = get_DivMod_mem(n);
1776 turn_into_tuple(n, 4);
1777 set_Tuple_pred(n, pn_DivMod_M, mem);
1778 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1779 set_Tuple_pred(n, pn_DivMod_res_div, a);
1780 set_Tuple_pred(n, pn_DivMod_res_mod, b);
1787 * Optimize Abs(x) into x if x is Confirmed >= 0
1788 * Optimize Abs(x) into -x if x is Confirmed <= 0
1790 static ir_node *transform_node_Abs(ir_node *n)
1793 ir_node *a = get_Abs_op(n);
1794 value_classify sign = classify_value_sign(a);
1796 if (sign == VALUE_NEGATIVE) {
1797 ir_mode *mode = get_irn_mode(n);
1800 * We can replace the Abs by -x here.
1801 * We even could add a new Confirm here.
1803 * Note that -x would create a new node, so we could
1804 * not run it in the equivalent_node() context.
1806 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
1807 get_irn_n(n, -1), a, mode);
1809 DBG_OPT_CONFIRM(oldn, n);
1811 else if (sign == VALUE_POSITIVE) {
1812 /* n is positive, Abs is not needed */
1815 DBG_OPT_CONFIRM(oldn, n);
1822 * transform a Cond node
1824 static ir_node *transform_node_Cond(ir_node *n)
1826 /* Replace the Cond by a Jmp if it branches on a constant
1829 ir_node *a = get_Cond_selector(n);
1830 tarval *ta = value_of(a);
1832 /* we need block info which is not available in floating irgs */
1833 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
1836 if ((ta != tarval_bad) &&
1837 (get_irn_mode(a) == mode_b) &&
1838 (get_opt_unreachable_code())) {
1839 /* It's a boolean Cond, branching on a boolean constant.
1840 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
1841 jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n));
1842 turn_into_tuple(n, 2);
1843 if (ta == tarval_b_true) {
1844 set_Tuple_pred(n, pn_Cond_false, new_Bad());
1845 set_Tuple_pred(n, pn_Cond_true, jmp);
1847 set_Tuple_pred(n, pn_Cond_false, jmp);
1848 set_Tuple_pred(n, pn_Cond_true, new_Bad());
1850 /* We might generate an endless loop, so keep it alive. */
1851 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
1859 static ir_node *transform_node_Eor(ir_node *n)
1862 ir_node *a = get_Eor_left(n);
1863 ir_node *b = get_Eor_right(n);
1864 ir_mode *mode = get_irn_mode(n);
1868 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
1869 mode, get_mode_null(mode));
1870 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
1872 else if ((mode == mode_b)
1873 && (get_irn_op(a) == op_Proj)
1874 && (get_irn_mode(a) == mode_b)
1875 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)
1876 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
1877 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
1878 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
1879 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
1881 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
1883 else if ((mode == mode_b)
1884 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) {
1885 /* The Eor is a Not. Replace it by a Not. */
1886 /* ????!!!Extend to bitfield 1111111. */
1887 n = new_r_Not(current_ir_graph, get_irn_n(n, -1), a, mode_b);
1889 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
1896 * Transform a boolean Not.
1898 static ir_node *transform_node_Not(ir_node *n)
1901 ir_node *a = get_Not_op(n);
1903 if ( (get_irn_mode(n) == mode_b)
1904 && (get_irn_op(a) == op_Proj)
1905 && (get_irn_mode(a) == mode_b)
1906 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
1907 /* We negate a Cmp. The Cmp has the negated result anyways! */
1908 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
1909 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
1910 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
1917 * Transform a Cast_type(Const) into a new Const_type
1919 static ir_node *transform_node_Cast(ir_node *n) {
1921 ir_node *pred = get_Cast_op(n);
1922 type *tp = get_irn_type(n);
1924 if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) {
1925 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
1926 get_Const_tarval(pred), tp);
1927 DBG_OPT_CSTEVAL(oldn, n);
1928 } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) {
1929 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred),
1930 get_SymConst_kind(pred), tp);
1931 DBG_OPT_CSTEVAL(oldn, n);
1938 * Transform a Proj(Div) with a non-zero value.
1939 * Removes the exceptions and routes the memory to the NoMem node.
1941 static ir_node *transform_node_Proj_Div(ir_node *proj)
1943 ir_node *n = get_Proj_pred(proj);
1944 ir_node *b = get_Div_right(n);
1947 if (value_not_zero(b)) {
1948 /* div(x, y) && y != 0 */
1949 proj_nr = get_Proj_proj(proj);
1951 /* this node may float */
1952 set_irn_pinned(n, op_pin_state_floats);
1954 if (proj_nr == pn_Div_X_except) {
1955 /* we found an exception handler, remove it */
1957 } else if (proj_nr == pn_Div_M) {
1958 /* the memory Proj can be removed */
1959 ir_node *res = get_Div_mem(n);
1960 set_Div_mem(n, get_irg_no_mem(current_ir_graph));
1969 * Transform a Proj(Mod) with a non-zero value.
1970 * Removes the exceptions and routes the memory to the NoMem node.
1972 static ir_node *transform_node_Proj_Mod(ir_node *proj)
1974 ir_node *n = get_Proj_pred(proj);
1975 ir_node *b = get_Mod_right(n);
1978 if (value_not_zero(b)) {
1979 /* mod(x, y) && y != 0 */
1980 proj_nr = get_Proj_proj(proj);
1982 /* this node may float */
1983 set_irn_pinned(n, op_pin_state_floats);
1985 if (proj_nr == pn_Mod_X_except) {
1986 /* we found an exception handler, remove it */
1988 } else if (proj_nr == pn_Mod_M) {
1989 /* the memory Proj can be removed */
1990 ir_node *res = get_Mod_mem(n);
1991 set_Mod_mem(n, get_irg_no_mem(current_ir_graph));
1995 else if (proj_nr == pn_Mod_res && get_Mod_left(n) == b) {
1996 /* a % a = 0 if a != 0 */
1997 ir_mode *mode = get_irn_mode(proj);
1998 ir_node *res = new_Const(mode, get_mode_null(mode));
2000 DBG_OPT_CSTEVAL(n, res);
2008 * Transform a Proj(DivMod) with a non-zero value.
2009 * Removes the exceptions and routes the memory to the NoMem node.
2011 static ir_node *transform_node_Proj_DivMod(ir_node *proj)
2013 ir_node *n = get_Proj_pred(proj);
2014 ir_node *b = get_DivMod_right(n);
2017 if (value_not_zero(b)) {
2018 /* DivMod(x, y) && y != 0 */
2019 proj_nr = get_Proj_proj(proj);
2021 /* this node may float */
2022 set_irn_pinned(n, op_pin_state_floats);
2024 if (proj_nr == pn_DivMod_X_except) {
2025 /* we found an exception handler, remove it */
2028 else if (proj_nr == pn_DivMod_M) {
2029 /* the memory Proj can be removed */
2030 ir_node *res = get_DivMod_mem(n);
2031 set_DivMod_mem(n, get_irg_no_mem(current_ir_graph));
2035 else if (proj_nr == pn_DivMod_res_mod && get_DivMod_left(n) == b) {
2036 /* a % a = 0 if a != 0 */
2037 ir_mode *mode = get_irn_mode(proj);
2038 ir_node *res = new_Const(mode, get_mode_null(mode));
2040 DBG_OPT_CSTEVAL(n, res);
2048 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
2050 static ir_node *transform_node_Proj_Cond(ir_node *proj)
2052 if (get_opt_unreachable_code()) {
2053 ir_node *n = get_Proj_pred(proj);
2054 ir_node *b = get_Cond_selector(n);
2056 if (mode_is_int(get_irn_mode(b))) {
2057 tarval *tb = value_of(b);
2059 if (tb != tarval_bad) {
2060 /* we have a constant switch */
2061 long num = get_Proj_proj(proj);
2063 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
2064 if (get_tarval_long(tb) == num) {
2065 /* Do NOT create a jump here, or we will have 2 control flow ops
2066 * in a block. This case is optimized away in optimize_cf(). */
2070 /* this case will NEVER be taken, kill it */
2081 * Normalizes and optimizes Cmp nodes.
2083 static ir_node *transform_node_Proj_Cmp(ir_node *proj)
2085 if (get_opt_reassociation()) {
2086 ir_node *n = get_Proj_pred(proj);
2087 ir_node *left = get_Cmp_left(n);
2088 ir_node *right = get_Cmp_right(n);
2092 ir_mode *mode = NULL;
2093 long proj_nr = get_Proj_proj(proj);
2096 * First step: normalize the compare op
2097 * by placing the constant on the right site
2098 * or moving the lower address node to the left.
2099 * We ignore the case that both are constants
2100 * this case should be optimized away.
2102 if (get_irn_op(right) == op_Const)
2104 else if (get_irn_op(left) == op_Const) {
2109 proj_nr = get_inversed_pnc(proj_nr);
2112 else if (left > right) {
2118 proj_nr = get_inversed_pnc(proj_nr);
2123 * Second step: Try to reduce the magnitude
2124 * of a constant. This may help to generate better code
2125 * later and may help to normalize more compares.
2126 * Of course this is only possible for integer values.
2129 mode = get_irn_mode(c);
2130 tv = get_Const_tarval(c);
2132 if (tv != tarval_bad) {
2133 /* the following optimization is possible on modes without Overflow
2134 * on Unary Minus or on == and !=:
2135 * -a CMP c ==> a swap(CMP) -c
2137 * Beware: for two-complement Overflow may occur, so only == and != can
2138 * be optimized, see this:
2139 * -MININT < 0 =/=> MININT > 0 !!!
2141 if (get_opt_constant_folding() && get_irn_op(left) == op_Minus &&
2142 (!mode_overflow_on_unary_Minus(mode) ||
2143 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
2144 left = get_Minus_op(left);
2145 tv = tarval_sub(get_mode_null(mode), tv);
2147 proj_nr = get_inversed_pnc(proj_nr);
2151 /* for integer modes, we have more */
2152 if (mode_is_int(mode)) {
2153 /* Ne includes Unordered which is not possible on integers.
2154 * However, frontends often use this wrong, so fix it here */
2155 if (proj_nr & pn_Cmp_Uo) {
2156 proj_nr &= ~pn_Cmp_Uo;
2157 set_Proj_proj(proj, proj_nr);
2160 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
2161 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
2162 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
2163 tv = tarval_sub(tv, get_mode_one(mode));
2165 proj_nr ^= pn_Cmp_Eq;
2168 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
2169 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
2170 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
2171 tv = tarval_add(tv, get_mode_one(mode));
2173 proj_nr ^= pn_Cmp_Eq;
2177 /* the following reassociations work only for == and != */
2178 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
2180 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
2181 if (classify_tarval(tv) == TV_CLASSIFY_NULL && get_irn_op(left) == op_Sub) {
2182 right = get_Sub_right(left);
2183 left = get_Sub_left(left);
2185 tv = value_of(right);
2189 if (tv != tarval_bad) {
2190 ir_op *op = get_irn_op(left);
2192 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
2194 ir_node *c1 = get_Sub_right(left);
2195 tarval *tv2 = value_of(c1);
2197 if (tv2 != tarval_bad) {
2198 tv2 = tarval_add(tv, value_of(c1));
2200 if (tv2 != tarval_bad) {
2201 left = get_Sub_left(left);
2207 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
2208 else if (op == op_Add) {
2209 ir_node *a_l = get_Add_left(left);
2210 ir_node *a_r = get_Add_right(left);
2214 if (get_irn_op(a_l) == op_Const) {
2216 tv2 = value_of(a_l);
2220 tv2 = value_of(a_r);
2223 if (tv2 != tarval_bad) {
2224 tv2 = tarval_sub(tv, tv2);
2226 if (tv2 != tarval_bad) {
2233 /* -a == c ==> a == -c, -a != c ==> a != -c */
2234 else if (op == op_Minus) {
2235 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
2237 if (tv2 != tarval_bad) {
2238 left = get_Minus_op(left);
2245 /* the following reassociations work only for <= */
2246 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2247 if (tv != tarval_bad) {
2248 ir_op *op = get_irn_op(left);
2250 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
2260 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
2262 if (changed & 2) /* need a new Const */
2263 right = new_Const(mode, tv);
2265 /* create a new compare */
2266 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block,
2269 set_Proj_pred(proj, n);
2270 set_Proj_proj(proj, proj_nr);
2277 * Does all optimizations on nodes that must be done on it's Proj's
2278 * because of creating new nodes.
2280 static ir_node *transform_node_Proj(ir_node *proj)
2282 ir_node *n = get_Proj_pred(proj);
2284 switch (get_irn_opcode(n)) {
2286 return transform_node_Proj_Div(proj);
2289 return transform_node_Proj_Mod(proj);
2292 return transform_node_Proj_DivMod(proj);
2295 return transform_node_Proj_Cond(proj);
2298 return transform_node_Proj_Cmp(proj);
2301 /* should not happen, but if it does will be optimized away */
2302 return equivalent_node_Proj(proj);
2311 * returns the operands of a commutative bin-op, if one operand is
2312 * a const, it is returned as the second one.
2314 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c)
2316 ir_node *op_a = get_binop_left(binop);
2317 ir_node *op_b = get_binop_right(binop);
2319 assert(is_op_commutative(get_irn_op(binop)));
2321 if (get_irn_op(op_a) == op_Const) {
2332 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
2333 * Such pattern may arise in bitfield stores.
2335 * value c4 value c4 & c2
2336 * AND c3 AND c1 | c3
2341 static ir_node *transform_node_Or_bf_store(ir_node *or)
2345 ir_node *and_l, *c3;
2346 ir_node *value, *c4;
2347 ir_node *new_and, *new_const, *block;
2348 ir_mode *mode = get_irn_mode(or);
2350 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
2352 get_comm_Binop_Ops(or, &and, &c1);
2353 if ((get_irn_op(c1) != op_Const) || (get_irn_op(and) != op_And))
2356 get_comm_Binop_Ops(and, &or_l, &c2);
2357 if ((get_irn_op(c2) != op_Const) || (get_irn_op(or_l) != op_Or))
2360 get_comm_Binop_Ops(or_l, &and_l, &c3);
2361 if ((get_irn_op(c3) != op_Const) || (get_irn_op(and_l) != op_And))
2364 get_comm_Binop_Ops(and_l, &value, &c4);
2365 if (get_irn_op(c4) != op_Const)
2368 /* ok, found the pattern, check for conditions */
2369 assert(mode == get_irn_mode(and));
2370 assert(mode == get_irn_mode(or_l));
2371 assert(mode == get_irn_mode(and_l));
2373 tv1 = get_Const_tarval(c1);
2374 tv2 = get_Const_tarval(c2);
2375 tv3 = get_Const_tarval(c3);
2376 tv4 = get_Const_tarval(c4);
2378 tv = tarval_or(tv4, tv2);
2379 if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) {
2380 /* have at least one 0 at the same bit position */
2384 n_tv4 = tarval_not(tv4);
2385 if (tv3 != tarval_and(tv3, n_tv4)) {
2386 /* bit in the or_mask is outside the and_mask */
2390 n_tv2 = tarval_not(tv2);
2391 if (tv1 != tarval_and(tv1, n_tv2)) {
2392 /* bit in the or_mask is outside the and_mask */
2396 /* ok, all conditions met */
2397 block = get_irn_n(or, -1);
2399 new_and = new_r_And(current_ir_graph, block,
2400 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
2402 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
2404 set_Or_left(or, new_and);
2405 set_Or_right(or, new_const);
2407 /* check for more */
2408 return transform_node_Or_bf_store(or);
2412 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
2414 static ir_node *transform_node_Or_Rot(ir_node *or)
2416 ir_mode *mode = get_irn_mode(or);
2417 ir_node *shl, *shr, *block;
2418 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
2421 if (! mode_is_int(mode))
2424 shl = get_binop_left(or);
2425 shr = get_binop_right(or);
2427 if (get_irn_op(shl) == op_Shr) {
2428 if (get_irn_op(shr) != op_Shl)
2435 else if (get_irn_op(shl) != op_Shl)
2437 else if (get_irn_op(shr) != op_Shr)
2440 x = get_Shl_left(shl);
2441 if (x != get_Shr_left(shr))
2444 c1 = get_Shl_right(shl);
2445 c2 = get_Shr_right(shr);
2446 if (get_irn_op(c1) == op_Const && get_irn_op(c2) == op_Const) {
2447 tv1 = get_Const_tarval(c1);
2448 if (! tarval_is_long(tv1))
2451 tv2 = get_Const_tarval(c2);
2452 if (! tarval_is_long(tv2))
2455 if (get_tarval_long(tv1) + get_tarval_long(tv2)
2456 != get_mode_size_bits(mode))
2459 /* yet, condition met */
2460 block = get_irn_n(or, -1);
2462 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
2464 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
2467 else if (get_irn_op(c1) == op_Sub) {
2471 if (get_Sub_right(sub) != v)
2474 c1 = get_Sub_left(sub);
2475 if (get_irn_op(c1) != op_Const)
2478 tv1 = get_Const_tarval(c1);
2479 if (! tarval_is_long(tv1))
2482 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
2485 /* yet, condition met */
2486 block = get_nodes_block(or);
2488 /* a Rot right is not supported, so use a rot left */
2489 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
2491 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
2494 else if (get_irn_op(c2) == op_Sub) {
2498 c1 = get_Sub_left(sub);
2499 if (get_irn_op(c1) != op_Const)
2502 tv1 = get_Const_tarval(c1);
2503 if (! tarval_is_long(tv1))
2506 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
2509 /* yet, condition met */
2510 block = get_irn_n(or, -1);
2513 n = new_r_Rot(current_ir_graph, block, x, v, mode);
2515 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
2525 static ir_node *transform_node_Or(ir_node *or)
2527 or = transform_node_Or_bf_store(or);
2528 or = transform_node_Or_Rot(or);
2534 static ir_node *transform_node(ir_node *n);
2537 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl.
2539 * Should be moved to reassociation?
2541 static ir_node *transform_node_shift(ir_node *n)
2543 ir_node *left, *right;
2544 tarval *tv1, *tv2, *res;
2546 int modulo_shf, flag;
2548 left = get_binop_left(n);
2550 /* different operations */
2551 if (get_irn_op(left) != get_irn_op(n))
2554 right = get_binop_right(n);
2555 tv1 = value_of(right);
2556 if (tv1 == tarval_bad)
2559 tv2 = value_of(get_binop_right(left));
2560 if (tv2 == tarval_bad)
2563 res = tarval_add(tv1, tv2);
2565 /* beware: a simple replacement works only, if res < modulo shift */
2566 mode = get_irn_mode(n);
2570 modulo_shf = get_mode_modulo_shift(mode);
2571 if (modulo_shf > 0) {
2572 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
2574 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
2581 /* ok, we can replace it */
2582 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
2584 in[0] = get_binop_left(left);
2585 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
2587 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
2589 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
2591 return transform_node(irn);
2596 #define transform_node_Shr transform_node_shift
2597 #define transform_node_Shrs transform_node_shift
2598 #define transform_node_Shl transform_node_shift
2601 * Remove dead blocks and nodes in dead blocks
2602 * in keep alive list. We do not generate a new End node.
2604 static ir_node *transform_node_End(ir_node *n) {
2605 int i, n_keepalives = get_End_n_keepalives(n);
2607 for (i = 0; i < n_keepalives; ++i) {
2608 ir_node *ka = get_End_keepalive(n, i);
2610 if (is_Block_dead(ka)) {
2611 set_End_keepalive(n, i, new_Bad());
2614 else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka)))
2615 set_End_keepalive(n, i, new_Bad());
2621 * Optimize a Mux into some simpler cases.
2623 static ir_node *transform_node_Mux(ir_node *n)
2625 ir_node *oldn = n, *sel = get_Mux_sel(n);
2626 ir_mode *mode = get_irn_mode(n);
2628 if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(mode)) {
2629 ir_node *cmp = get_Proj_pred(sel);
2630 long proj_nr = get_Proj_proj(sel);
2631 ir_node *f = get_Mux_false(n);
2632 ir_node *t = get_Mux_true(n);
2634 if (get_irn_op(cmp) == op_Cmp && classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
2635 ir_node *block = get_irn_n(n, -1);
2638 * Note: normalization puts the constant on the right site,
2639 * so we check only one case.
2641 * Note further that these optimization work even for floating point
2642 * with NaN's because -NaN == NaN.
2643 * However, if +0 and -0 is handled differently, we cannot use the first one.
2645 if (get_irn_op(f) == op_Minus &&
2646 get_Minus_op(f) == t &&
2647 get_Cmp_left(cmp) == t) {
2649 if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
2650 /* Mux(a >=/> 0, -a, a) ==> Abs(a) */
2651 n = new_rd_Abs(get_irn_dbg_info(n),
2655 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2658 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2659 /* Mux(a <=/< 0, -a, a) ==> Minus(Abs(a)) */
2660 n = new_rd_Abs(get_irn_dbg_info(n),
2664 n = new_rd_Minus(get_irn_dbg_info(n),
2669 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2673 else if (get_irn_op(t) == op_Minus &&
2674 get_Minus_op(t) == f &&
2675 get_Cmp_left(cmp) == f) {
2677 if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2678 /* Mux(a <=/< 0, a, -a) ==> Abs(a) */
2679 n = new_rd_Abs(get_irn_dbg_info(n),
2683 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2686 else if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
2687 /* Mux(a >=/> 0, a, -a) ==> Minus(Abs(a)) */
2688 n = new_rd_Abs(get_irn_dbg_info(n),
2692 n = new_rd_Minus(get_irn_dbg_info(n),
2697 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2702 if (mode_is_int(mode) && mode_is_signed(mode) &&
2703 get_mode_arithmetic(mode) == irma_twos_complement) {
2704 ir_node *x = get_Cmp_left(cmp);
2706 /* the following optimization works only with signed integer two-complement mode */
2708 if (mode == get_irn_mode(x)) {
2710 * FIXME: this restriction is two rigid, as it would still
2711 * work if mode(x) = Hs and mode == Is, but at least it removes
2714 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Le) &&
2715 classify_Const(t) == CNST_ALL_ONE &&
2716 classify_Const(f) == CNST_NULL) {
2718 * Mux(x:T </<= 0, 0, -1) -> Shrs(x, sizeof_bits(T) - 1)
2722 n = new_rd_Shrs(get_irn_dbg_info(n),
2723 current_ir_graph, block, x,
2724 new_r_Const_long(current_ir_graph, block, mode_Iu,
2725 get_mode_size_bits(mode) - 1),
2727 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
2730 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) &&
2731 classify_Const(t) == CNST_ONE &&
2732 classify_Const(f) == CNST_NULL) {
2734 * Mux(x:T >/>= 0, 0, 1) -> Shr(-x, sizeof_bits(T) - 1)
2738 n = new_rd_Shr(get_irn_dbg_info(n),
2739 current_ir_graph, block,
2740 new_r_Minus(current_ir_graph, block, x, mode),
2741 new_r_Const_long(current_ir_graph, block, mode_Iu,
2742 get_mode_size_bits(mode) - 1),
2744 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
2751 return arch_transform_node_Mux(n);
2755 * Tries several [inplace] [optimizing] transformations and returns an
2756 * equivalent node. The difference to equivalent_node() is that these
2757 * transformations _do_ generate new nodes, and thus the old node must
2758 * not be freed even if the equivalent node isn't the old one.
2760 static ir_node *transform_node(ir_node *n)
2762 if (n->op->transform_node)
2763 n = n->op->transform_node(n);
2768 * set the default transform node operation
2770 static ir_op *firm_set_default_transform_node(ir_op *op)
2774 op->transform_node = transform_node_##a; \
2798 op->transform_node = NULL;
2806 /* **************** Common Subexpression Elimination **************** */
2808 /** The size of the hash table used, should estimate the number of nodes
2810 #define N_IR_NODES 512
2812 /** Compares the attributes of two Const nodes. */
2813 static int node_cmp_attr_Const(ir_node *a, ir_node *b)
2815 return (get_Const_tarval(a) != get_Const_tarval(b))
2816 || (get_Const_type(a) != get_Const_type(b));
2819 /** Compares the attributes of two Proj nodes. */
2820 static int node_cmp_attr_Proj(ir_node *a, ir_node *b)
2822 return get_irn_proj_attr (a) != get_irn_proj_attr (b);
2825 /** Compares the attributes of two Filter nodes. */
2826 static int node_cmp_attr_Filter(ir_node *a, ir_node *b)
2828 return get_Filter_proj(a) != get_Filter_proj(b);
2831 /** Compares the attributes of two Alloc nodes. */
2832 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b)
2834 return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where)
2835 || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type);
2838 /** Compares the attributes of two Free nodes. */
2839 static int node_cmp_attr_Free(ir_node *a, ir_node *b)
2841 return (get_irn_free_attr(a).where != get_irn_free_attr(b).where)
2842 || (get_irn_free_attr(a).type != get_irn_free_attr(b).type);
2845 /** Compares the attributes of two SymConst nodes. */
2846 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b)
2848 return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num)
2849 || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p)
2850 || (get_irn_symconst_attr(a).tp != get_irn_symconst_attr(b).tp);
2853 /** Compares the attributes of two Call nodes. */
2854 static int node_cmp_attr_Call(ir_node *a, ir_node *b)
2856 return (get_irn_call_attr(a) != get_irn_call_attr(b));
2859 /** Compares the attributes of two Sel nodes. */
2860 static int node_cmp_attr_Sel(ir_node *a, ir_node *b)
2862 return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind)
2863 || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name)
2864 || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner)
2865 || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name)
2866 || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type);
2869 /** Compares the attributes of two Phi nodes. */
2870 static int node_cmp_attr_Phi(ir_node *a, ir_node *b)
2872 return get_irn_phi_attr (a) != get_irn_phi_attr (b);
2875 /** Compares the attributes of two Cast nodes. */
2876 static int node_cmp_attr_Cast(ir_node *a, ir_node *b)
2878 return get_Cast_type(a) != get_Cast_type(b);
2881 /** Compares the attributes of two Load nodes. */
2882 static int node_cmp_attr_Load(ir_node *a, ir_node *b)
2884 if (get_Load_volatility(a) == volatility_is_volatile ||
2885 get_Load_volatility(b) == volatility_is_volatile)
2886 /* NEVER do CSE on volatile Loads */
2889 return get_Load_mode(a) != get_Load_mode(b);
2892 /** Compares the attributes of two Store nodes. */
2893 static int node_cmp_attr_Store(ir_node *a, ir_node *b)
2895 /* NEVER do CSE on volatile Stores */
2896 return (get_Store_volatility(a) == volatility_is_volatile ||
2897 get_Store_volatility(b) == volatility_is_volatile);
2900 /** Compares the attributes of two Confirm nodes. */
2901 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b)
2903 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
2907 * set the default node attribute compare operation
2909 static ir_op *firm_set_default_node_cmp_attr(ir_op *op)
2913 op->node_cmp_attr = node_cmp_attr_##a; \
2931 op->node_cmp_attr = NULL;
2939 * Compare function for two nodes in the hash table. Gets two
2940 * nodes as parameters. Returns 0 if the nodes are a cse.
2943 vt_cmp (const void *elt, const void *key)
2951 if (a == b) return 0;
2953 if ((get_irn_op(a) != get_irn_op(b)) ||
2954 (get_irn_mode(a) != get_irn_mode(b))) return 1;
2956 /* compare if a's in and b's in are of equal length */
2957 irn_arity_a = get_irn_intra_arity (a);
2958 if (irn_arity_a != get_irn_intra_arity(b))
2961 /* for block-local cse and op_pin_state_pinned nodes: */
2962 if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) {
2963 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
2967 /* compare a->in[0..ins] with b->in[0..ins] */
2968 for (i = 0; i < irn_arity_a; i++)
2969 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
2973 * here, we already now that the nodes are identical except their
2976 if (a->op->node_cmp_attr)
2977 return a->op->node_cmp_attr(a, b);
2983 * Calculate a hash value of a node.
2986 ir_node_hash (ir_node *node)
2991 if (node->op == op_Const) {
2992 /* special value for const, as they only differ in their tarval. */
2993 h = HASH_PTR(node->attr.con.tv);
2994 h = 9*h + HASH_PTR(get_irn_mode(node));
2995 } else if (node->op == op_SymConst) {
2996 /* special value for const, as they only differ in their symbol. */
2997 h = HASH_PTR(node->attr.i.sym.type_p);
2998 h = 9*h + HASH_PTR(get_irn_mode(node));
3001 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
3002 h = irn_arity = get_irn_intra_arity(node);
3004 /* consider all in nodes... except the block if not a control flow. */
3005 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
3006 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
3010 h = 9*h + HASH_PTR(get_irn_mode(node));
3012 h = 9*h + HASH_PTR(get_irn_op(node));
3019 new_identities(void) {
3020 return new_pset(vt_cmp, N_IR_NODES);
3024 del_identities(pset *value_table) {
3025 del_pset(value_table);
3029 * Return the canonical node computing the same value as n.
3030 * Looks up the node in a hash table.
3032 * For Const nodes this is performed in the constructor, too. Const
3033 * nodes are extremely time critical because of their frequent use in
3034 * constant string arrays.
3036 static INLINE ir_node *
3037 identify (pset *value_table, ir_node *n)
3041 if (!value_table) return n;
3043 if (get_opt_reassociation()) {
3044 if (is_op_commutative(get_irn_op(n))) {
3045 ir_node *l = get_binop_left(n);
3046 ir_node *r = get_binop_right(n);
3048 /* for commutative operators perform a OP b == b OP a */
3050 set_binop_left(n, r);
3051 set_binop_right(n, l);
3056 o = pset_find (value_table, n, ir_node_hash (n));
3065 * During construction we set the op_pin_state_pinned flag in the graph right when the
3066 * optimization is performed. The flag turning on procedure global cse could
3067 * be changed between two allocations. This way we are safe.
3069 static INLINE ir_node *
3070 identify_cons (pset *value_table, ir_node *n) {
3073 n = identify(value_table, n);
3074 if (get_irn_n(old, -1) != get_irn_n(n, -1))
3075 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3080 * Return the canonical node computing the same value as n.
3081 * Looks up the node in a hash table, enters it in the table
3082 * if it isn't there yet.
3085 identify_remember (pset *value_table, ir_node *n)
3089 if (!value_table) return n;
3091 if (get_opt_reassociation()) {
3092 if (is_op_commutative(get_irn_op(n))) {
3093 ir_node *l = get_binop_left(n);
3094 ir_node *r = get_binop_right(n);
3096 /* for commutative operators perform a OP b == b OP a */
3098 set_binop_left(n, r);
3099 set_binop_right(n, l);
3104 /* lookup or insert in hash table with given hash key. */
3105 o = pset_insert (value_table, n, ir_node_hash (n));
3115 add_identities (pset *value_table, ir_node *node) {
3116 if (get_opt_cse() && (get_irn_opcode(node) != iro_Block))
3117 identify_remember (value_table, node);
3121 * garbage in, garbage out. If a node has a dead input, i.e., the
3122 * Bad node is input to the node, return the Bad node.
3124 static INLINE ir_node *
3125 gigo (ir_node *node)
3128 ir_op *op = get_irn_op(node);
3130 /* remove garbage blocks by looking at control flow that leaves the block
3131 and replacing the control flow by Bad. */
3132 if (get_irn_mode(node) == mode_X) {
3133 ir_node *block = get_nodes_block(skip_Proj(node));
3135 /* Don't optimize nodes in immature blocks. */
3136 if (!get_Block_matured(block)) return node;
3137 /* Don't optimize End, may have Bads. */
3138 if (op == op_End) return node;
3140 if (is_Block(block)) {
3141 irn_arity = get_irn_arity(block);
3142 for (i = 0; i < irn_arity; i++) {
3143 if (!is_Bad(get_irn_n(block, i)))
3146 if (i == irn_arity) return new_Bad();
3150 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
3151 blocks predecessors is dead. */
3152 if ( op != op_Block && op != op_Phi && op != op_Tuple) {
3153 irn_arity = get_irn_arity(node);
3156 * Beware: we can only read the block of a non-floating node.
3158 if (is_irn_pinned_in_irg(node) &&
3159 is_Block_dead(get_nodes_block(node)))
3162 for (i = 0; i < irn_arity; i++) {
3163 ir_node *pred = get_irn_n(node, i);
3167 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
3168 return new_Unknown(get_irn_mode(node));
3172 /* With this code we violate the agreement that local_optimize
3173 only leaves Bads in Block, Phi and Tuple nodes. */
3174 /* If Block has only Bads as predecessors it's garbage. */
3175 /* If Phi has only Bads as predecessors it's garbage. */
3176 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
3177 irn_arity = get_irn_arity(node);
3178 for (i = 0; i < irn_arity; i++) {
3179 if (!is_Bad(get_irn_n(node, i))) break;
3181 if (i == irn_arity) node = new_Bad();
3189 * These optimizations deallocate nodes from the obstack.
3190 * It can only be called if it is guaranteed that no other nodes
3191 * reference this one, i.e., right after construction of a node.
3193 * current_ir_graph must be set to the graph of the node!
3196 optimize_node(ir_node *n)
3200 opcode iro = get_irn_opcode(n);
3202 /* Always optimize Phi nodes: part of the construction. */
3203 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
3205 /* constant expression evaluation / constant folding */
3206 if (get_opt_constant_folding()) {
3207 /* neither constants nor Tuple values can be evaluated */
3208 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
3209 /* try to evaluate */
3210 tv = computed_value(n);
3211 if (tv != tarval_bad) {
3213 type *old_tp = get_irn_type(n);
3214 int i, arity = get_irn_arity(n);
3218 * Try to recover the type of the new expression.
3220 for (i = 0; i < arity && !old_tp; ++i)
3221 old_tp = get_irn_type(get_irn_n(n, i));
3224 * we MUST copy the node here temporary, because it's still needed
3225 * for DBG_OPT_CSTEVAL
3227 node_size = offsetof(ir_node, attr) + n->op->attr_size;
3228 oldn = alloca(node_size);
3230 memcpy(oldn, n, node_size);
3231 CLONE_ARR_A(ir_node *, oldn->in, n->in);
3233 /* ARG, copy the in array, we need it for statistics */
3234 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
3236 /* note the inplace edges module */
3237 edges_node_deleted(n, current_ir_graph);
3239 /* evaluation was successful -- replace the node. */
3240 obstack_free(current_ir_graph->obst, n);
3241 nw = new_Const(get_tarval_mode (tv), tv);
3243 if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv))
3244 set_Const_type(nw, old_tp);
3245 DBG_OPT_CSTEVAL(oldn, nw);
3251 /* remove unnecessary nodes */
3252 if (get_opt_constant_folding() ||
3253 (iro == iro_Phi) || /* always optimize these nodes. */
3255 (iro == iro_Proj) ||
3256 (iro == iro_Block) ) /* Flags tested local. */
3257 n = equivalent_node (n);
3259 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
3261 /* Common Subexpression Elimination.
3263 * Checks whether n is already available.
3264 * The block input is used to distinguish different subexpressions. Right
3265 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
3266 * subexpressions within a block.
3269 n = identify_cons (current_ir_graph->value_table, n);
3272 edges_node_deleted(oldn, current_ir_graph);
3274 /* We found an existing, better node, so we can deallocate the old node. */
3275 obstack_free (current_ir_graph->obst, oldn);
3280 /* Some more constant expression evaluation that does not allow to
3282 iro = get_irn_opcode(n);
3283 if (get_opt_constant_folding() ||
3284 (iro == iro_Cond) ||
3285 (iro == iro_Proj) ||
3286 (iro == iro_Sel)) /* Flags tested local. */
3287 n = transform_node (n);
3289 /* Remove nodes with dead (Bad) input.
3290 Run always for transformation induced Bads. */
3293 /* Now we have a legal, useful node. Enter it in hash table for CSE */
3294 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
3295 n = identify_remember (current_ir_graph->value_table, n);
3303 * These optimizations never deallocate nodes (in place). This can cause dead
3304 * nodes lying on the obstack. Remove these by a dead node elimination,
3305 * i.e., a copying garbage collection.
3308 optimize_in_place_2 (ir_node *n)
3312 opcode iro = get_irn_opcode(n);
3314 if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n;
3316 /* constant expression evaluation / constant folding */
3317 if (get_opt_constant_folding()) {
3318 /* neither constants nor Tuple values can be evaluated */
3319 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
3320 /* try to evaluate */
3321 tv = computed_value(n);
3322 if (tv != tarval_bad) {
3323 /* evaluation was successful -- replace the node. */
3324 type *old_tp = get_irn_type(n);
3325 int i, arity = get_irn_arity(n);
3328 * Try to recover the type of the new expression.
3330 for (i = 0; i < arity && !old_tp; ++i)
3331 old_tp = get_irn_type(get_irn_n(n, i));
3333 n = new_Const(get_tarval_mode(tv), tv);
3335 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
3336 set_Const_type(n, old_tp);
3338 DBG_OPT_CSTEVAL(oldn, n);
3344 /* remove unnecessary nodes */
3345 if (get_opt_constant_folding() ||
3346 (iro == iro_Phi) || /* always optimize these nodes. */
3347 (iro == iro_Id) || /* ... */
3348 (iro == iro_Proj) || /* ... */
3349 (iro == iro_Block) ) /* Flags tested local. */
3350 n = equivalent_node(n);
3352 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
3354 /** common subexpression elimination **/
3355 /* Checks whether n is already available. */
3356 /* The block input is used to distinguish different subexpressions. Right
3357 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
3358 subexpressions within a block. */
3359 if (get_opt_cse()) {
3360 n = identify(current_ir_graph->value_table, n);
3363 /* Some more constant expression evaluation. */
3364 iro = get_irn_opcode(n);
3365 if (get_opt_constant_folding() ||
3366 (iro == iro_Cond) ||
3367 (iro == iro_Proj) ||
3368 (iro == iro_Sel)) /* Flags tested local. */
3369 n = transform_node(n);
3371 /* Remove nodes with dead (Bad) input.
3372 Run always for transformation induced Bads. */
3375 /* Now we can verify the node, as it has no dead inputs any more. */
3378 /* Now we have a legal, useful node. Enter it in hash table for cse.
3379 Blocks should be unique anyways. (Except the successor of start:
3380 is cse with the start block!) */
3381 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
3382 n = identify_remember(current_ir_graph->value_table, n);
3388 * Wrapper for external use, set proper status bits after optimization.
3391 optimize_in_place (ir_node *n)
3393 /* Handle graph state */
3394 assert(get_irg_phase_state(current_ir_graph) != phase_building);
3396 if (get_opt_global_cse())
3397 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3398 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
3399 set_irg_outs_inconsistent(current_ir_graph);
3401 /* Maybe we could also test whether optimizing the node can
3402 change the control graph. */
3403 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
3404 set_irg_dom_inconsistent(current_ir_graph);
3405 return optimize_in_place_2 (n);
3409 * set the default ir op operations
3411 ir_op *firm_set_default_operations(ir_op *op)
3413 op = firm_set_default_computed_value(op);
3414 op = firm_set_default_equivalent_node(op);
3415 op = firm_set_default_transform_node(op);
3416 op = firm_set_default_node_cmp_attr(op);
3417 op = firm_set_default_get_type(op);