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);
760 * Returns a equivalent node for a Jmp, a Bad :-)
761 * Of course this only happens if the Block of the Jmp is Bad.
763 static ir_node *equivalent_node_Jmp(ir_node *n)
765 /* unreachable code elimination */
766 if (is_Block_dead(get_nodes_block(n)))
772 /* Same for op_Raise */
773 #define equivalent_node_Raise equivalent_node_Jmp
776 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
777 See transform_node_Proj_Cond(). */
780 * optimize operations that are commutative and have neutral 0,
781 * so a op 0 = 0 op a = a.
783 static ir_node *equivalent_node_neutral_zero(ir_node *n)
787 ir_node *a = get_binop_left(n);
788 ir_node *b = get_binop_right(n);
793 /* After running compute_node there is only one constant predecessor.
794 Find this predecessors value and remember the other node: */
795 if ((tv = value_of(a)) != tarval_bad) {
797 } else if ((tv = value_of(b)) != tarval_bad) {
802 /* If this predecessors constant value is zero, the operation is
803 * unnecessary. Remove it.
805 * Beware: If n is a Add, the mode of on and n might be different
806 * which happens in this rare construction: NULL + 3.
807 * Then, a Conv would be needed which we cannot include here.
809 if (classify_tarval (tv) == TV_CLASSIFY_NULL) {
810 if (get_irn_mode(on) == get_irn_mode(n)) {
813 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
820 #define equivalent_node_Eor equivalent_node_neutral_zero
823 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
825 * The second one looks strange, but this construct
826 * is used heavily in the LCC sources :-).
828 * Beware: The Mode of an Add may be different than the mode of its
829 * predecessors, so we could not return a predecessors in all cases.
831 static ir_node *equivalent_node_Add(ir_node *n)
834 ir_node *left, *right;
836 n = equivalent_node_neutral_zero(n);
840 left = get_Add_left(n);
841 right = get_Add_right(n);
843 if (get_irn_op(left) == op_Sub) {
844 if (get_Sub_right(left) == right) {
847 n = get_Sub_left(left);
848 if (get_irn_mode(oldn) == get_irn_mode(n)) {
849 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
854 if (get_irn_op(right) == op_Sub) {
855 if (get_Sub_right(right) == left) {
858 n = get_Sub_left(right);
859 if (get_irn_mode(oldn) == get_irn_mode(n)) {
860 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
869 * optimize operations that are not commutative but have neutral 0 on left,
872 static ir_node *equivalent_node_left_zero(ir_node *n)
876 ir_node *a = get_binop_left(n);
877 ir_node *b = get_binop_right(n);
879 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
882 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
888 #define equivalent_node_Shl equivalent_node_left_zero
889 #define equivalent_node_Shr equivalent_node_left_zero
890 #define equivalent_node_Shrs equivalent_node_left_zero
891 #define equivalent_node_Rot equivalent_node_left_zero
894 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
896 * The second one looks strange, but this construct
897 * is used heavily in the LCC sources :-).
899 * Beware: The Mode of a Sub may be different than the mode of its
900 * predecessors, so we could not return a predecessors in all cases.
902 static ir_node *equivalent_node_Sub(ir_node *n)
906 ir_node *a = get_Sub_left(n);
907 ir_node *b = get_Sub_right(n);
909 /* Beware: modes might be different */
910 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
911 if (get_irn_mode(n) == get_irn_mode(a)) {
914 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
917 else if (get_irn_op(a) == op_Add) {
918 ir_mode *mode = get_irn_mode(n);
920 if (mode_wrap_around(mode)) {
921 ir_node *left = get_Add_left(a);
922 ir_node *right = get_Add_right(a);
925 if (get_irn_mode(n) == get_irn_mode(right)) {
927 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
930 else if (right == b) {
931 if (get_irn_mode(n) == get_irn_mode(left)) {
933 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
944 * Optimize an "idempotent unary op", ie op(op(n)) = n.
946 * @fixme -(-a) == a, but might overflow two times.
947 * We handle it anyway here but the better way would be a
948 * flag. This would be needed for Pascal for instance.
950 static ir_node *equivalent_node_idempotent_unop(ir_node *n)
953 ir_node *pred = get_unop_op(n);
955 /* optimize symmetric unop */
956 if (get_irn_op(pred) == get_irn_op(n)) {
957 n = get_unop_op(pred);
958 DBG_OPT_ALGSIM2(oldn, pred, n);
963 /* Not(Not(x)) == x */
964 #define equivalent_node_Not equivalent_node_idempotent_unop
966 /* --x == x */ /* ??? Is this possible or can --x raise an
967 out of bounds exception if min =! max? */
968 #define equivalent_node_Minus equivalent_node_idempotent_unop
971 * Optimize a * 1 = 1 * a = a.
973 static ir_node *equivalent_node_Mul(ir_node *n)
977 ir_node *a = get_Mul_left(n);
978 ir_node *b = get_Mul_right(n);
980 /* Mul is commutative and has again an other neutral element. */
981 if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) {
983 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
984 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) {
986 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
992 * Optimize a / 1 = a.
994 static ir_node *equivalent_node_Div(ir_node *n)
996 ir_node *a = get_Div_left(n);
997 ir_node *b = get_Div_right(n);
999 /* Div is not commutative. */
1000 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1001 /* Turn Div into a tuple (mem, bad, a) */
1002 ir_node *mem = get_Div_mem(n);
1003 turn_into_tuple(n, 3);
1004 set_Tuple_pred(n, pn_Div_M, mem);
1005 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
1006 set_Tuple_pred(n, pn_Div_res, a);
1012 * Optimize a / 1 = a.
1014 static ir_node *equivalent_node_DivMod(ir_node *n)
1016 ir_node *a = get_DivMod_left(n);
1017 ir_node *b = get_DivMod_right(n);
1019 /* Div is not commutative. */
1020 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1021 /* Turn DivMod into a tuple (mem, bad, a, 0) */
1022 ir_node *mem = get_Div_mem(n);
1023 ir_mode *mode = get_irn_mode(b);
1025 turn_into_tuple(n, 4);
1026 set_Tuple_pred(n, pn_DivMod_M, mem);
1027 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1028 set_Tuple_pred(n, pn_DivMod_res_div, a);
1029 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1035 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1037 static ir_node *equivalent_node_Or(ir_node *n)
1041 ir_node *a = get_Or_left(n);
1042 ir_node *b = get_Or_right(n);
1045 n = a; /* Or has it's own neutral element */
1046 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1047 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) {
1049 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1050 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
1052 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1059 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a.
1061 static ir_node *equivalent_node_And(ir_node *n)
1065 ir_node *a = get_And_left(n);
1066 ir_node *b = get_And_right(n);
1069 n = a; /* And has it's own neutral element */
1070 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1071 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) {
1073 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1074 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) {
1076 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1082 * Try to remove useless Conv's:
1084 static ir_node *equivalent_node_Conv(ir_node *n)
1087 ir_node *a = get_Conv_op(n);
1090 ir_mode *n_mode = get_irn_mode(n);
1091 ir_mode *a_mode = get_irn_mode(a);
1093 if (n_mode == a_mode) { /* No Conv necessary */
1095 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1096 } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */
1100 n_mode = get_irn_mode(n);
1101 b_mode = get_irn_mode(b);
1103 if (n_mode == b_mode) {
1104 if (n_mode == mode_b) {
1105 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1106 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1108 else if (mode_is_int(n_mode) || mode_is_character(n_mode)) {
1109 if (smaller_mode(b_mode, a_mode)){
1110 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1111 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1120 * A Cast may be removed if the type of the previous node
1121 * is already the type of the Cast.
1123 static ir_node *equivalent_node_Cast(ir_node *n) {
1125 ir_node *pred = get_Cast_op(n);
1127 if (get_irn_type(pred) == get_Cast_type(n)) {
1129 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1134 /* Several optimizations:
1135 - no Phi in start block.
1136 - remove Id operators that are inputs to Phi
1137 - fold Phi-nodes, iff they have only one predecessor except
1140 static ir_node *equivalent_node_Phi(ir_node *n)
1145 ir_node *block = NULL; /* to shutup gcc */
1146 ir_node *first_val = NULL; /* to shutup gcc */
1147 ir_node *scnd_val = NULL; /* to shutup gcc */
1149 if (!get_opt_normalize()) return n;
1151 n_preds = get_Phi_n_preds(n);
1153 block = get_nodes_block(n);
1154 /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
1155 assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
1156 if ((is_Block_dead(block)) || /* Control dead */
1157 (block == current_ir_graph->start_block)) /* There should be no Phi nodes */
1158 return new_Bad(); /* in the Start Block. */
1160 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1162 /* If the Block has a Bad pred, we also have one. */
1163 for (i = 0; i < n_preds; ++i)
1164 if (is_Bad(get_Block_cfgpred(block, i)))
1165 set_Phi_pred(n, i, new_Bad());
1167 /* Find first non-self-referencing input */
1168 for (i = 0; i < n_preds; ++i) {
1169 first_val = get_Phi_pred(n, i);
1170 if ( (first_val != n) /* not self pointer */
1172 && (! is_Bad(first_val))
1174 ) { /* value not dead */
1175 break; /* then found first value. */
1180 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1186 /* follow_Id () for rest of inputs, determine if any of these
1187 are non-self-referencing */
1188 while (++i < n_preds) {
1189 scnd_val = get_Phi_pred(n, i);
1190 if ( (scnd_val != n)
1191 && (scnd_val != first_val)
1193 && (! is_Bad(scnd_val))
1201 /* Fold, if no multiple distinct non-self-referencing inputs */
1203 DBG_OPT_PHI(oldn, n);
1205 /* skip the remaining Ids (done in get_Phi_pred). */
1206 /* superfluous, since we walk all to propagate Block's Bads.
1207 while (++i < n_preds) get_Phi_pred(n, i); */
1213 * optimize Proj(Tuple) and gigo() for ProjX in Bad block
1215 static ir_node *equivalent_node_Proj(ir_node *n)
1219 ir_node *a = get_Proj_pred(n);
1221 if ( get_irn_op(a) == op_Tuple) {
1222 /* Remove the Tuple/Proj combination. */
1223 if ( get_Proj_proj(n) <= get_Tuple_n_preds(a) ) {
1224 n = get_Tuple_pred(a, get_Proj_proj(n));
1225 DBG_OPT_TUPLE(oldn, a, n);
1227 assert(0); /* This should not happen! */
1231 else if (get_irn_mode(n) == mode_X) {
1232 if (is_Block_dead(get_nodes_block(skip_Proj(n)))) {
1233 /* Remove dead control flow -- early gigo(). */
1244 static ir_node *equivalent_node_Id(ir_node *n)
1250 } while (get_irn_op(n) == op_Id);
1252 DBG_OPT_ID(oldn, n);
1259 static ir_node *equivalent_node_Mux(ir_node *n)
1261 ir_node *oldn = n, *sel = get_Mux_sel(n);
1262 tarval *ts = value_of(sel);
1264 /* Mux(true, f, t) == t */
1265 if (ts == tarval_b_true) {
1266 n = get_Mux_true(n);
1267 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1269 /* Mux(false, f, t) == f */
1270 else if (ts == tarval_b_false) {
1271 n = get_Mux_false(n);
1272 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1274 /* Mux(v, x, x) == x */
1275 else if (get_Mux_false(n) == get_Mux_true(n)) {
1276 n = get_Mux_true(n);
1277 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1279 else if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(get_irn_mode(n))) {
1280 ir_node *cmp = get_Proj_pred(sel);
1281 long proj_nr = get_Proj_proj(sel);
1282 ir_node *b = get_Mux_false(n);
1283 ir_node *a = get_Mux_true(n);
1286 * Note: normalization puts the constant on the right site,
1287 * so we check only one case.
1289 * Note further that these optimization work even for floating point
1290 * with NaN's because -NaN == NaN.
1291 * However, if +0 and -0 is handled differently, we cannot use the first one.
1293 if (get_irn_op(cmp) == op_Cmp && get_Cmp_left(cmp) == a) {
1294 if (classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
1295 /* Mux(a CMP 0, X, a) */
1296 if (get_irn_op(b) == op_Minus && get_Minus_op(b) == a) {
1297 /* Mux(a CMP 0, -a, a) */
1298 if (proj_nr == pn_Cmp_Eq) {
1299 /* Mux(a == 0, -a, a) ==> -a */
1301 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1303 else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1304 /* Mux(a != 0, -a, a) ==> a */
1306 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1309 else if (classify_Const(b) == CNST_NULL) {
1310 /* Mux(a CMP 0, 0, a) */
1311 if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1312 /* Mux(a != 0, 0, a) ==> a */
1314 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1316 else if (proj_nr == pn_Cmp_Eq) {
1317 /* Mux(a == 0, 0, a) ==> 0 */
1319 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1329 * Optimize -a CMP -b into b CMP a.
1330 * This works only for for modes where unary Minus
1332 * Note that two-complement integers can Overflow
1333 * so it will NOT work.
1335 static ir_node *equivalent_node_Cmp(ir_node *n)
1337 ir_node *left = get_Cmp_left(n);
1338 ir_node *right = get_Cmp_right(n);
1340 if (get_irn_op(left) == op_Minus && get_irn_op(right) == op_Minus &&
1341 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1342 left = get_Minus_op(left);
1343 right = get_Minus_op(right);
1344 set_Cmp_left(n, right);
1345 set_Cmp_right(n, left);
1351 * Remove Confirm nodes if setting is on.
1352 * Replace Confirms(x, '=', Constlike) by Constlike.
1354 static ir_node *equivalent_node_Confirm(ir_node *n)
1356 ir_node *pred = get_Confirm_value(n);
1357 pn_Cmp pnc = get_Confirm_cmp(n);
1359 if (get_irn_op(pred) == op_Confirm && pnc == get_Confirm_cmp(pred)) {
1361 * rare case: two identical Confirms one after another,
1362 * replace the second one with the first.
1366 if (pnc == pn_Cmp_Eq) {
1367 ir_node *bound = get_Confirm_bound(n);
1370 * Optimize a rare case:
1371 * Confirm(x, '=', Constlike) ==> Constlike
1373 if (is_irn_constlike(bound)) {
1374 DBG_OPT_CONFIRM(n, bound);
1378 return get_opt_remove_Confirm() ? get_Confirm_value(n) : n;
1382 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1383 * perform no actual computation, as, e.g., the Id nodes. It does not create
1384 * new nodes. It is therefore safe to free n if the node returned is not n.
1385 * If a node returns a Tuple we can not just skip it. If the size of the
1386 * in array fits, we transform n into a tuple (e.g., Div).
1389 equivalent_node(ir_node *n)
1391 if (n->op->equivalent_node)
1392 return n->op->equivalent_node(n);
1397 * set the default equivalent node operation
1399 static ir_op *firm_set_default_equivalent_node(ir_op *op)
1403 op->equivalent_node = equivalent_node_##a; \
1433 op->equivalent_node = NULL;
1441 * Do node specific optimizations of nodes predecessors.
1444 optimize_preds(ir_node *n) {
1445 ir_node *a = NULL, *b = NULL;
1447 /* get the operands we will work on for simple cases. */
1449 a = get_binop_left(n);
1450 b = get_binop_right(n);
1451 } else if (is_unop(n)) {
1455 switch (get_irn_opcode(n)) {
1458 /* We don't want Cast as input to Cmp. */
1459 if (get_irn_op(a) == op_Cast) {
1463 if (get_irn_op(b) == op_Cast) {
1465 set_Cmp_right(n, b);
1474 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1475 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1476 * If possible, remove the Conv's.
1478 static ir_node *transform_node_AddSub(ir_node *n)
1480 ir_mode *mode = get_irn_mode(n);
1482 if (mode_is_reference(mode)) {
1483 ir_node *left = get_binop_left(n);
1484 ir_node *right = get_binop_right(n);
1485 int ref_bits = get_mode_size_bits(mode);
1487 if (get_irn_op(left) == op_Conv) {
1488 ir_mode *mode = get_irn_mode(left);
1489 int bits = get_mode_size_bits(mode);
1491 if (ref_bits == bits &&
1492 mode_is_int(mode) &&
1493 get_mode_arithmetic(mode) == irma_twos_complement) {
1494 ir_node *pre = get_Conv_op(left);
1495 ir_mode *pre_mode = get_irn_mode(pre);
1497 if (mode_is_int(pre_mode) &&
1498 get_mode_size_bits(pre_mode) == bits &&
1499 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1500 /* ok, this conv just changes to sign, moreover the calculation
1501 * is done with same number of bits as our address mode, so
1502 * we can ignore the conv as address calculation can be viewed
1503 * as either signed or unsigned
1505 set_binop_left(n, pre);
1510 if (get_irn_op(right) == op_Conv) {
1511 ir_mode *mode = get_irn_mode(right);
1512 int bits = get_mode_size_bits(mode);
1514 if (ref_bits == bits &&
1515 mode_is_int(mode) &&
1516 get_mode_arithmetic(mode) == irma_twos_complement) {
1517 ir_node *pre = get_Conv_op(right);
1518 ir_mode *pre_mode = get_irn_mode(pre);
1520 if (mode_is_int(pre_mode) &&
1521 get_mode_size_bits(pre_mode) == bits &&
1522 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1523 /* ok, this conv just changes to sign, moreover the calculation
1524 * is done with same number of bits as our address mode, so
1525 * we can ignore the conv as address calculation can be viewed
1526 * as either signed or unsigned
1528 set_binop_right(n, pre);
1537 * Do the AddSub optimization, then Transform Add(a,a) into Mul(a, 2)
1538 * if the mode is integer or float.
1539 * Transform Add(a,-b) into Sub(a,b).
1540 * Reassociation might fold this further.
1542 static ir_node *transform_node_Add(ir_node *n)
1547 n = transform_node_AddSub(n);
1549 mode = get_irn_mode(n);
1550 if (mode_is_num(mode)) {
1551 ir_node *a = get_Add_left(n);
1553 if (a == get_Add_right(n)) {
1554 ir_node *block = get_irn_n(n, -1);
1557 get_irn_dbg_info(n),
1561 new_r_Const_long(current_ir_graph, block, mode, 2),
1563 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
1566 ir_node *b = get_Add_right(n);
1568 if (get_irn_op(a) == op_Minus) {
1570 get_irn_dbg_info(n),
1576 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1578 else if (get_irn_op(b) == op_Minus) {
1580 get_irn_dbg_info(n),
1586 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1594 * Do the AddSub optimization, then Transform Sub(0,a) into Minus(a).
1596 static ir_node *transform_node_Sub(ir_node *n)
1601 n = transform_node_AddSub(n);
1603 mode = get_irn_mode(n);
1604 if (mode_is_num(mode) && (classify_Const(get_Sub_left(n)) == CNST_NULL)) {
1606 get_irn_dbg_info(n),
1611 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
1618 * Transform Mul(a,-1) into -a.
1619 * Do architecture dependent optimizations on Mul nodes
1621 static ir_node *transform_node_Mul(ir_node *n) {
1623 ir_mode *mode = get_irn_mode(n);
1625 if (mode_is_signed(mode)) {
1627 ir_node *a = get_Mul_left(n);
1628 ir_node *b = get_Mul_right(n);
1630 if (value_of(a) == get_mode_minus_one(mode))
1632 else if (value_of(b) == get_mode_minus_one(mode))
1635 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
1636 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
1640 return arch_dep_replace_mul_with_shifts(n);
1644 * transform a Div Node
1646 static ir_node *transform_node_Div(ir_node *n)
1648 tarval *tv = value_of(n);
1651 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
1653 if (tv != tarval_bad) {
1654 value = new_Const(get_tarval_mode(tv), tv);
1656 DBG_OPT_CSTEVAL(n, value);
1658 else /* Try architecture dependent optimization */
1659 value = arch_dep_replace_div_by_const(n);
1662 /* Turn Div into a tuple (mem, bad, value) */
1663 ir_node *mem = get_Div_mem(n);
1665 turn_into_tuple(n, pn_Div_max);
1666 set_Tuple_pred(n, pn_Div_M, mem);
1667 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
1668 set_Tuple_pred(n, pn_Div_res, value);
1674 * transform a Mod node
1676 static ir_node *transform_node_Mod(ir_node *n)
1678 tarval *tv = value_of(n);
1681 /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */
1683 if (tv != tarval_bad) {
1684 value = new_Const(get_tarval_mode(tv), tv);
1686 DBG_OPT_CSTEVAL(n, value);
1688 else /* Try architecture dependent optimization */
1689 value = arch_dep_replace_mod_by_const(n);
1692 /* Turn Mod into a tuple (mem, bad, value) */
1693 ir_node *mem = get_Mod_mem(n);
1695 turn_into_tuple(n, 3);
1696 set_Tuple_pred(n, pn_Mod_M, mem);
1697 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
1698 set_Tuple_pred(n, pn_Mod_res, value);
1704 * transform a DivMod node
1706 static ir_node *transform_node_DivMod(ir_node *n)
1710 ir_node *a = get_DivMod_left(n);
1711 ir_node *b = get_DivMod_right(n);
1712 ir_mode *mode = get_irn_mode(a);
1713 tarval *ta = value_of(a);
1714 tarval *tb = value_of(b);
1716 if (!(mode_is_int(mode) && mode_is_int(get_irn_mode(b))))
1719 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
1721 if (tb != tarval_bad) {
1722 if (tb == get_mode_one(get_tarval_mode(tb))) {
1723 b = new_Const (mode, get_mode_null(mode));
1726 DBG_OPT_CSTEVAL(n, b);
1728 else if (ta != tarval_bad) {
1729 tarval *resa, *resb;
1730 resa = tarval_div (ta, tb);
1731 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
1732 Jmp for X result!? */
1733 resb = tarval_mod (ta, tb);
1734 if (resb == tarval_bad) return n; /* Causes exception! */
1735 a = new_Const (mode, resa);
1736 b = new_Const (mode, resb);
1739 DBG_OPT_CSTEVAL(n, a);
1740 DBG_OPT_CSTEVAL(n, b);
1742 else { /* Try architecture dependent optimization */
1743 arch_dep_replace_divmod_by_const(&a, &b, n);
1744 evaluated = a != NULL;
1746 } else if (ta == get_mode_null(mode)) {
1747 /* 0 / non-Const = 0 */
1752 if (evaluated) { /* replace by tuple */
1753 ir_node *mem = get_DivMod_mem(n);
1754 turn_into_tuple(n, 4);
1755 set_Tuple_pred(n, pn_DivMod_M, mem);
1756 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1757 set_Tuple_pred(n, pn_DivMod_res_div, a);
1758 set_Tuple_pred(n, pn_DivMod_res_mod, b);
1765 * Optimize Abs(x) into x if x is Confirmed >= 0
1766 * Optimize Abs(x) into -x if x is Confirmed <= 0
1768 static ir_node *transform_node_Abs(ir_node *n)
1771 ir_node *a = get_Abs_op(n);
1772 value_classify sign = classify_value_sign(a);
1774 if (sign == VALUE_NEGATIVE) {
1775 ir_mode *mode = get_irn_mode(n);
1778 * We can replace the Abs by -x here.
1779 * We even could add a new Confirm here.
1781 * Note that -x would create a new node, so we could
1782 * not run it in the equivalent_node() context.
1784 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
1785 get_irn_n(n, -1), a, mode);
1787 DBG_OPT_CONFIRM(oldn, n);
1789 else if (sign == VALUE_POSITIVE) {
1790 /* n is positive, Abs is not needed */
1793 DBG_OPT_CONFIRM(oldn, n);
1800 * transform a Cond node
1802 static ir_node *transform_node_Cond(ir_node *n)
1804 /* Replace the Cond by a Jmp if it branches on a constant
1807 ir_node *a = get_Cond_selector(n);
1808 tarval *ta = value_of(a);
1810 /* we need block info which is not available in floating irgs */
1811 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
1814 if ((ta != tarval_bad) &&
1815 (get_irn_mode(a) == mode_b) &&
1816 (get_opt_unreachable_code())) {
1817 /* It's a boolean Cond, branching on a boolean constant.
1818 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
1819 jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n));
1820 turn_into_tuple(n, 2);
1821 if (ta == tarval_b_true) {
1822 set_Tuple_pred(n, pn_Cond_false, new_Bad());
1823 set_Tuple_pred(n, pn_Cond_true, jmp);
1825 set_Tuple_pred(n, pn_Cond_false, jmp);
1826 set_Tuple_pred(n, pn_Cond_true, new_Bad());
1828 /* We might generate an endless loop, so keep it alive. */
1829 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
1837 static ir_node *transform_node_Eor(ir_node *n)
1840 ir_node *a = get_Eor_left(n);
1841 ir_node *b = get_Eor_right(n);
1842 ir_mode *mode = get_irn_mode(n);
1846 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
1847 mode, get_mode_null(mode));
1848 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
1850 else if ((mode == mode_b)
1851 && (get_irn_op(a) == op_Proj)
1852 && (get_irn_mode(a) == mode_b)
1853 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)
1854 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
1855 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
1856 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
1857 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
1859 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
1861 else if ((mode == mode_b)
1862 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) {
1863 /* The Eor is a Not. Replace it by a Not. */
1864 /* ????!!!Extend to bitfield 1111111. */
1865 n = new_r_Not(current_ir_graph, get_irn_n(n, -1), a, mode_b);
1867 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
1874 * Transform a boolean Not.
1876 static ir_node *transform_node_Not(ir_node *n)
1879 ir_node *a = get_Not_op(n);
1881 if ( (get_irn_mode(n) == mode_b)
1882 && (get_irn_op(a) == op_Proj)
1883 && (get_irn_mode(a) == mode_b)
1884 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
1885 /* We negate a Cmp. The Cmp has the negated result anyways! */
1886 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
1887 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
1888 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
1895 * Transform a Cast_type(Const) into a new Const_type
1897 static ir_node *transform_node_Cast(ir_node *n) {
1899 ir_node *pred = get_Cast_op(n);
1900 type *tp = get_irn_type(n);
1902 if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) {
1903 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
1904 get_Const_tarval(pred), tp);
1905 DBG_OPT_CSTEVAL(oldn, n);
1906 } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) {
1907 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred),
1908 get_SymConst_kind(pred), tp);
1909 DBG_OPT_CSTEVAL(oldn, n);
1916 * Transform a Proj(Div) with a non-zero value.
1917 * Removes the exceptions and routes the memory to the NoMem node.
1919 static ir_node *transform_node_Proj_Div(ir_node *proj)
1921 ir_node *n = get_Proj_pred(proj);
1922 ir_node *b = get_Div_right(n);
1925 if (value_not_zero(b)) {
1926 /* div(x, y) && y != 0 */
1927 proj_nr = get_Proj_proj(proj);
1929 /* this node may float */
1930 set_irn_pinned(n, op_pin_state_floats);
1932 if (proj_nr == pn_Div_X_except) {
1933 /* we found an exception handler, remove it */
1935 } else if (proj_nr == pn_Div_M) {
1936 /* the memory Proj can be removed */
1937 ir_node *res = get_Div_mem(n);
1938 set_Div_mem(n, get_irg_no_mem(current_ir_graph));
1947 * Transform a Proj(Mod) with a non-zero value.
1948 * Removes the exceptions and routes the memory to the NoMem node.
1950 static ir_node *transform_node_Proj_Mod(ir_node *proj)
1952 ir_node *n = get_Proj_pred(proj);
1953 ir_node *b = get_Mod_right(n);
1956 if (value_not_zero(b)) {
1957 /* mod(x, y) && y != 0 */
1958 proj_nr = get_Proj_proj(proj);
1960 /* this node may float */
1961 set_irn_pinned(n, op_pin_state_floats);
1963 if (proj_nr == pn_Mod_X_except) {
1964 /* we found an exception handler, remove it */
1966 } else if (proj_nr == pn_Mod_M) {
1967 /* the memory Proj can be removed */
1968 ir_node *res = get_Mod_mem(n);
1969 set_Mod_mem(n, get_irg_no_mem(current_ir_graph));
1973 else if (proj_nr == pn_Mod_res && get_Mod_left(n) == b) {
1974 /* a % a = 0 if a != 0 */
1975 ir_mode *mode = get_irn_mode(proj);
1976 ir_node *res = new_Const(mode, get_mode_null(mode));
1978 DBG_OPT_CSTEVAL(n, res);
1986 * Transform a Proj(DivMod) with a non-zero value.
1987 * Removes the exceptions and routes the memory to the NoMem node.
1989 static ir_node *transform_node_Proj_DivMod(ir_node *proj)
1991 ir_node *n = get_Proj_pred(proj);
1992 ir_node *b = get_DivMod_right(n);
1995 if (value_not_zero(b)) {
1996 /* DivMod(x, y) && y != 0 */
1997 proj_nr = get_Proj_proj(proj);
1999 /* this node may float */
2000 set_irn_pinned(n, op_pin_state_floats);
2002 if (proj_nr == pn_DivMod_X_except) {
2003 /* we found an exception handler, remove it */
2006 else if (proj_nr == pn_DivMod_M) {
2007 /* the memory Proj can be removed */
2008 ir_node *res = get_DivMod_mem(n);
2009 set_DivMod_mem(n, get_irg_no_mem(current_ir_graph));
2013 else if (proj_nr == pn_DivMod_res_mod && get_DivMod_left(n) == b) {
2014 /* a % a = 0 if a != 0 */
2015 ir_mode *mode = get_irn_mode(proj);
2016 ir_node *res = new_Const(mode, get_mode_null(mode));
2018 DBG_OPT_CSTEVAL(n, res);
2026 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
2028 static ir_node *transform_node_Proj_Cond(ir_node *proj)
2030 if (get_opt_unreachable_code()) {
2031 ir_node *n = get_Proj_pred(proj);
2032 ir_node *b = get_Cond_selector(n);
2034 if (mode_is_int(get_irn_mode(b))) {
2035 tarval *tb = value_of(b);
2037 if (tb != tarval_bad) {
2038 /* we have a constant switch */
2039 long num = get_Proj_proj(proj);
2041 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
2042 if (get_tarval_long(tb) == num) {
2043 /* Do NOT create a jump here, or we will have 2 control flow ops
2044 * in a block. This case is optimized away in optimize_cf(). */
2048 /* this case will NEVER be taken, kill it */
2059 * Normalizes and optimizes Cmp nodes.
2061 static ir_node *transform_node_Proj_Cmp(ir_node *proj)
2063 if (get_opt_reassociation()) {
2064 ir_node *n = get_Proj_pred(proj);
2065 ir_node *left = get_Cmp_left(n);
2066 ir_node *right = get_Cmp_right(n);
2070 ir_mode *mode = NULL;
2071 long proj_nr = get_Proj_proj(proj);
2074 * First step: normalize the compare op
2075 * by placing the constant on the right site
2076 * or moving the lower address node to the left.
2077 * We ignore the case that both are constants
2078 * this case should be optimized away.
2080 if (get_irn_op(right) == op_Const)
2082 else if (get_irn_op(left) == op_Const) {
2087 proj_nr = get_inversed_pnc(proj_nr);
2090 else if (left > right) {
2096 proj_nr = get_inversed_pnc(proj_nr);
2101 * Second step: Try to reduce the magnitude
2102 * of a constant. This may help to generate better code
2103 * later and may help to normalize more compares.
2104 * Of course this is only possible for integer values.
2107 mode = get_irn_mode(c);
2108 tv = get_Const_tarval(c);
2110 if (tv != tarval_bad) {
2111 /* the following optimization is possible on modes without Overflow
2112 * on Unary Minus or on == and !=:
2113 * -a CMP c ==> a swap(CMP) -c
2115 * Beware: for two-complement Overflow may occur, so only == and != can
2116 * be optimized, see this:
2117 * -MININT < 0 =/=> MININT > 0 !!!
2119 if (get_opt_constant_folding() && get_irn_op(left) == op_Minus &&
2120 (!mode_overflow_on_unary_Minus(mode) ||
2121 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
2122 left = get_Minus_op(left);
2123 tv = tarval_sub(get_mode_null(mode), tv);
2125 proj_nr = get_inversed_pnc(proj_nr);
2129 /* for integer modes, we have more */
2130 if (mode_is_int(mode)) {
2131 /* Ne includes Unordered which is not possible on integers.
2132 * However, frontends often use this wrong, so fix it here */
2133 if (proj_nr & pn_Cmp_Uo) {
2134 proj_nr &= ~pn_Cmp_Uo;
2135 set_Proj_proj(proj, proj_nr);
2138 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
2139 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
2140 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
2141 tv = tarval_sub(tv, get_mode_one(mode));
2143 proj_nr ^= pn_Cmp_Eq;
2146 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
2147 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
2148 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
2149 tv = tarval_add(tv, get_mode_one(mode));
2151 proj_nr ^= pn_Cmp_Eq;
2155 /* the following reassociations work only for == and != */
2156 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
2158 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
2159 if (classify_tarval(tv) == TV_CLASSIFY_NULL && get_irn_op(left) == op_Sub) {
2160 right = get_Sub_right(left);
2161 left = get_Sub_left(left);
2163 tv = value_of(right);
2167 if (tv != tarval_bad) {
2168 ir_op *op = get_irn_op(left);
2170 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
2172 ir_node *c1 = get_Sub_right(left);
2173 tarval *tv2 = value_of(c1);
2175 if (tv2 != tarval_bad) {
2176 tv2 = tarval_add(tv, value_of(c1));
2178 if (tv2 != tarval_bad) {
2179 left = get_Sub_left(left);
2185 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
2186 else if (op == op_Add) {
2187 ir_node *a_l = get_Add_left(left);
2188 ir_node *a_r = get_Add_right(left);
2192 if (get_irn_op(a_l) == op_Const) {
2194 tv2 = value_of(a_l);
2198 tv2 = value_of(a_r);
2201 if (tv2 != tarval_bad) {
2202 tv2 = tarval_sub(tv, tv2);
2204 if (tv2 != tarval_bad) {
2211 /* -a == c ==> a == -c, -a != c ==> a != -c */
2212 else if (op == op_Minus) {
2213 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
2215 if (tv2 != tarval_bad) {
2216 left = get_Minus_op(left);
2223 /* the following reassociations work only for <= */
2224 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2225 if (tv != tarval_bad) {
2226 ir_op *op = get_irn_op(left);
2228 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
2238 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
2240 if (changed & 2) /* need a new Const */
2241 right = new_Const(mode, tv);
2243 /* create a new compare */
2244 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block,
2247 set_Proj_pred(proj, n);
2248 set_Proj_proj(proj, proj_nr);
2255 * Does all optimizations on nodes that must be done on it's Proj's
2256 * because of creating new nodes.
2258 static ir_node *transform_node_Proj(ir_node *proj)
2260 ir_node *n = get_Proj_pred(proj);
2262 switch (get_irn_opcode(n)) {
2264 return transform_node_Proj_Div(proj);
2267 return transform_node_Proj_Mod(proj);
2270 return transform_node_Proj_DivMod(proj);
2273 return transform_node_Proj_Cond(proj);
2276 return transform_node_Proj_Cmp(proj);
2279 /* should not happen, but if it does will be optimized away */
2280 return equivalent_node_Proj(proj);
2289 * returns the operands of a commutative bin-op, if one operand is
2290 * a const, it is returned as the second one.
2292 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c)
2294 ir_node *op_a = get_binop_left(binop);
2295 ir_node *op_b = get_binop_right(binop);
2297 assert(is_op_commutative(get_irn_op(binop)));
2299 if (get_irn_op(op_a) == op_Const) {
2310 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
2311 * Such pattern may arise in bitfield stores.
2313 * value c4 value c4 & c2
2314 * AND c3 AND c1 | c3
2319 static ir_node *transform_node_Or_bf_store(ir_node *or)
2323 ir_node *and_l, *c3;
2324 ir_node *value, *c4;
2325 ir_node *new_and, *new_const, *block;
2326 ir_mode *mode = get_irn_mode(or);
2328 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
2330 get_comm_Binop_Ops(or, &and, &c1);
2331 if ((get_irn_op(c1) != op_Const) || (get_irn_op(and) != op_And))
2334 get_comm_Binop_Ops(and, &or_l, &c2);
2335 if ((get_irn_op(c2) != op_Const) || (get_irn_op(or_l) != op_Or))
2338 get_comm_Binop_Ops(or_l, &and_l, &c3);
2339 if ((get_irn_op(c3) != op_Const) || (get_irn_op(and_l) != op_And))
2342 get_comm_Binop_Ops(and_l, &value, &c4);
2343 if (get_irn_op(c4) != op_Const)
2346 /* ok, found the pattern, check for conditions */
2347 assert(mode == get_irn_mode(and));
2348 assert(mode == get_irn_mode(or_l));
2349 assert(mode == get_irn_mode(and_l));
2351 tv1 = get_Const_tarval(c1);
2352 tv2 = get_Const_tarval(c2);
2353 tv3 = get_Const_tarval(c3);
2354 tv4 = get_Const_tarval(c4);
2356 tv = tarval_or(tv4, tv2);
2357 if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) {
2358 /* have at least one 0 at the same bit position */
2362 n_tv4 = tarval_not(tv4);
2363 if (tv3 != tarval_and(tv3, n_tv4)) {
2364 /* bit in the or_mask is outside the and_mask */
2368 n_tv2 = tarval_not(tv2);
2369 if (tv1 != tarval_and(tv1, n_tv2)) {
2370 /* bit in the or_mask is outside the and_mask */
2374 /* ok, all conditions met */
2375 block = get_irn_n(or, -1);
2377 new_and = new_r_And(current_ir_graph, block,
2378 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
2380 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
2382 set_Or_left(or, new_and);
2383 set_Or_right(or, new_const);
2385 /* check for more */
2386 return transform_node_Or_bf_store(or);
2390 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
2392 static ir_node *transform_node_Or_Rot(ir_node *or)
2394 ir_mode *mode = get_irn_mode(or);
2395 ir_node *shl, *shr, *block;
2396 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
2399 if (! mode_is_int(mode))
2402 shl = get_binop_left(or);
2403 shr = get_binop_right(or);
2405 if (get_irn_op(shl) == op_Shr) {
2406 if (get_irn_op(shr) != op_Shl)
2413 else if (get_irn_op(shl) != op_Shl)
2415 else if (get_irn_op(shr) != op_Shr)
2418 x = get_Shl_left(shl);
2419 if (x != get_Shr_left(shr))
2422 c1 = get_Shl_right(shl);
2423 c2 = get_Shr_right(shr);
2424 if (get_irn_op(c1) == op_Const && get_irn_op(c2) == op_Const) {
2425 tv1 = get_Const_tarval(c1);
2426 if (! tarval_is_long(tv1))
2429 tv2 = get_Const_tarval(c2);
2430 if (! tarval_is_long(tv2))
2433 if (get_tarval_long(tv1) + get_tarval_long(tv2)
2434 != get_mode_size_bits(mode))
2437 /* yet, condition met */
2438 block = get_irn_n(or, -1);
2440 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
2442 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
2445 else if (get_irn_op(c1) == op_Sub) {
2449 if (get_Sub_right(sub) != v)
2452 c1 = get_Sub_left(sub);
2453 if (get_irn_op(c1) != op_Const)
2456 tv1 = get_Const_tarval(c1);
2457 if (! tarval_is_long(tv1))
2460 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
2463 /* yet, condition met */
2464 block = get_nodes_block(or);
2466 /* a Rot right is not supported, so use a rot left */
2467 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
2469 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
2472 else if (get_irn_op(c2) == op_Sub) {
2476 c1 = get_Sub_left(sub);
2477 if (get_irn_op(c1) != op_Const)
2480 tv1 = get_Const_tarval(c1);
2481 if (! tarval_is_long(tv1))
2484 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
2487 /* yet, condition met */
2488 block = get_irn_n(or, -1);
2491 n = new_r_Rot(current_ir_graph, block, x, v, mode);
2493 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
2503 static ir_node *transform_node_Or(ir_node *or)
2505 or = transform_node_Or_bf_store(or);
2506 or = transform_node_Or_Rot(or);
2512 static ir_node *transform_node(ir_node *n);
2515 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl.
2517 * Should be moved to reassociation?
2519 static ir_node *transform_node_shift(ir_node *n)
2521 ir_node *left, *right;
2522 tarval *tv1, *tv2, *res;
2524 int modulo_shf, flag;
2526 left = get_binop_left(n);
2528 /* different operations */
2529 if (get_irn_op(left) != get_irn_op(n))
2532 right = get_binop_right(n);
2533 tv1 = value_of(right);
2534 if (tv1 == tarval_bad)
2537 tv2 = value_of(get_binop_right(left));
2538 if (tv2 == tarval_bad)
2541 res = tarval_add(tv1, tv2);
2543 /* beware: a simple replacement works only, if res < modulo shift */
2544 mode = get_irn_mode(n);
2548 modulo_shf = get_mode_modulo_shift(mode);
2549 if (modulo_shf > 0) {
2550 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
2552 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
2559 /* ok, we can replace it */
2560 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
2562 in[0] = get_binop_left(left);
2563 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
2565 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
2567 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
2569 return transform_node(irn);
2574 #define transform_node_Shr transform_node_shift
2575 #define transform_node_Shrs transform_node_shift
2576 #define transform_node_Shl transform_node_shift
2579 * Remove dead blocks and nodes in dead blocks
2580 * in keep alive list. We do not generate a new End node.
2582 static ir_node *transform_node_End(ir_node *n) {
2583 int i, n_keepalives = get_End_n_keepalives(n);
2585 for (i = 0; i < n_keepalives; ++i) {
2586 ir_node *ka = get_End_keepalive(n, i);
2588 if (is_Block_dead(ka)) {
2589 set_End_keepalive(n, i, new_Bad());
2592 else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka)))
2593 set_End_keepalive(n, i, new_Bad());
2599 * Optimize a Mux into some simpler cases.
2601 static ir_node *transform_node_Mux(ir_node *n)
2603 ir_node *oldn = n, *sel = get_Mux_sel(n);
2604 ir_mode *mode = get_irn_mode(n);
2606 if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(mode)) {
2607 ir_node *cmp = get_Proj_pred(sel);
2608 long proj_nr = get_Proj_proj(sel);
2609 ir_node *f = get_Mux_false(n);
2610 ir_node *t = get_Mux_true(n);
2612 if (get_irn_op(cmp) == op_Cmp && classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
2613 ir_node *block = get_irn_n(n, -1);
2616 * Note: normalization puts the constant on the right site,
2617 * so we check only one case.
2619 * Note further that these optimization work even for floating point
2620 * with NaN's because -NaN == NaN.
2621 * However, if +0 and -0 is handled differently, we cannot use the first one.
2623 if (get_irn_op(f) == op_Minus &&
2624 get_Minus_op(f) == t &&
2625 get_Cmp_left(cmp) == t) {
2627 if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
2628 /* Mux(a >=/> 0, -a, a) ==> Abs(a) */
2629 n = new_rd_Abs(get_irn_dbg_info(n),
2633 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2636 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2637 /* Mux(a <=/< 0, -a, a) ==> Minus(Abs(a)) */
2638 n = new_rd_Abs(get_irn_dbg_info(n),
2642 n = new_rd_Minus(get_irn_dbg_info(n),
2647 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2651 else if (get_irn_op(t) == op_Minus &&
2652 get_Minus_op(t) == f &&
2653 get_Cmp_left(cmp) == f) {
2655 if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2656 /* Mux(a <=/< 0, a, -a) ==> Abs(a) */
2657 n = new_rd_Abs(get_irn_dbg_info(n),
2661 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2664 else if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
2665 /* Mux(a >=/> 0, a, -a) ==> Minus(Abs(a)) */
2666 n = new_rd_Abs(get_irn_dbg_info(n),
2670 n = new_rd_Minus(get_irn_dbg_info(n),
2675 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
2680 if (mode_is_int(mode) && mode_is_signed(mode) &&
2681 get_mode_arithmetic(mode) == irma_twos_complement) {
2682 ir_node *x = get_Cmp_left(cmp);
2684 /* the following optimization works only with signed integer two-complement mode */
2686 if (mode == get_irn_mode(x)) {
2688 * FIXME: this restriction is two rigid, as it would still
2689 * work if mode(x) = Hs and mode == Is, but at least it removes
2692 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Le) &&
2693 classify_Const(t) == CNST_ALL_ONE &&
2694 classify_Const(f) == CNST_NULL) {
2696 * Mux(x:T </<= 0, 0, -1) -> Shrs(x, sizeof_bits(T) - 1)
2700 n = new_rd_Shrs(get_irn_dbg_info(n),
2701 current_ir_graph, block, x,
2702 new_r_Const_long(current_ir_graph, block, mode_Iu,
2703 get_mode_size_bits(mode) - 1),
2705 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
2708 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) &&
2709 classify_Const(t) == CNST_ONE &&
2710 classify_Const(f) == CNST_NULL) {
2712 * Mux(x:T >/>= 0, 0, 1) -> Shr(-x, sizeof_bits(T) - 1)
2716 n = new_rd_Shr(get_irn_dbg_info(n),
2717 current_ir_graph, block,
2718 new_r_Minus(current_ir_graph, block, x, mode),
2719 new_r_Const_long(current_ir_graph, block, mode_Iu,
2720 get_mode_size_bits(mode) - 1),
2722 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
2729 return arch_transform_node_Mux(n);
2733 * Tries several [inplace] [optimizing] transformations and returns an
2734 * equivalent node. The difference to equivalent_node() is that these
2735 * transformations _do_ generate new nodes, and thus the old node must
2736 * not be freed even if the equivalent node isn't the old one.
2738 static ir_node *transform_node(ir_node *n)
2740 if (n->op->transform_node)
2741 n = n->op->transform_node(n);
2746 * set the default transform node operation
2748 static ir_op *firm_set_default_transform_node(ir_op *op)
2752 op->transform_node = transform_node_##a; \
2776 op->transform_node = NULL;
2784 /* **************** Common Subexpression Elimination **************** */
2786 /** The size of the hash table used, should estimate the number of nodes
2788 #define N_IR_NODES 512
2790 /** Compares the attributes of two Const nodes. */
2791 static int node_cmp_attr_Const(ir_node *a, ir_node *b)
2793 return (get_Const_tarval(a) != get_Const_tarval(b))
2794 || (get_Const_type(a) != get_Const_type(b));
2797 /** Compares the attributes of two Proj nodes. */
2798 static int node_cmp_attr_Proj(ir_node *a, ir_node *b)
2800 return get_irn_proj_attr (a) != get_irn_proj_attr (b);
2803 /** Compares the attributes of two Filter nodes. */
2804 static int node_cmp_attr_Filter(ir_node *a, ir_node *b)
2806 return get_Filter_proj(a) != get_Filter_proj(b);
2809 /** Compares the attributes of two Alloc nodes. */
2810 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b)
2812 return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where)
2813 || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type);
2816 /** Compares the attributes of two Free nodes. */
2817 static int node_cmp_attr_Free(ir_node *a, ir_node *b)
2819 return (get_irn_free_attr(a).where != get_irn_free_attr(b).where)
2820 || (get_irn_free_attr(a).type != get_irn_free_attr(b).type);
2823 /** Compares the attributes of two SymConst nodes. */
2824 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b)
2826 return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num)
2827 || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p)
2828 || (get_irn_symconst_attr(a).tp != get_irn_symconst_attr(b).tp);
2831 /** Compares the attributes of two Call nodes. */
2832 static int node_cmp_attr_Call(ir_node *a, ir_node *b)
2834 return (get_irn_call_attr(a) != get_irn_call_attr(b));
2837 /** Compares the attributes of two Sel nodes. */
2838 static int node_cmp_attr_Sel(ir_node *a, ir_node *b)
2840 return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind)
2841 || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name)
2842 || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner)
2843 || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name)
2844 || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type);
2847 /** Compares the attributes of two Phi nodes. */
2848 static int node_cmp_attr_Phi(ir_node *a, ir_node *b)
2850 return get_irn_phi_attr (a) != get_irn_phi_attr (b);
2853 /** Compares the attributes of two Cast nodes. */
2854 static int node_cmp_attr_Cast(ir_node *a, ir_node *b)
2856 return get_Cast_type(a) != get_Cast_type(b);
2859 /** Compares the attributes of two Load nodes. */
2860 static int node_cmp_attr_Load(ir_node *a, ir_node *b)
2862 if (get_Load_volatility(a) == volatility_is_volatile ||
2863 get_Load_volatility(b) == volatility_is_volatile)
2864 /* NEVER do CSE on volatile Loads */
2867 return get_Load_mode(a) != get_Load_mode(b);
2870 /** Compares the attributes of two Store nodes. */
2871 static int node_cmp_attr_Store(ir_node *a, ir_node *b)
2873 /* NEVER do CSE on volatile Stores */
2874 return (get_Store_volatility(a) == volatility_is_volatile ||
2875 get_Store_volatility(b) == volatility_is_volatile);
2878 /** Compares the attributes of two Confirm nodes. */
2879 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b)
2881 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
2885 * set the default node attribute compare operation
2887 static ir_op *firm_set_default_node_cmp_attr(ir_op *op)
2891 op->node_cmp_attr = node_cmp_attr_##a; \
2909 op->node_cmp_attr = NULL;
2917 * Compare function for two nodes in the hash table. Gets two
2918 * nodes as parameters. Returns 0 if the nodes are a cse.
2921 vt_cmp (const void *elt, const void *key)
2929 if (a == b) return 0;
2931 if ((get_irn_op(a) != get_irn_op(b)) ||
2932 (get_irn_mode(a) != get_irn_mode(b))) return 1;
2934 /* compare if a's in and b's in are of equal length */
2935 irn_arity_a = get_irn_intra_arity (a);
2936 if (irn_arity_a != get_irn_intra_arity(b))
2939 /* for block-local cse and op_pin_state_pinned nodes: */
2940 if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) {
2941 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
2945 /* compare a->in[0..ins] with b->in[0..ins] */
2946 for (i = 0; i < irn_arity_a; i++)
2947 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
2951 * here, we already now that the nodes are identical except their
2954 if (a->op->node_cmp_attr)
2955 return a->op->node_cmp_attr(a, b);
2961 * Calculate a hash value of a node.
2964 ir_node_hash (ir_node *node)
2969 if (node->op == op_Const) {
2970 /* special value for const, as they only differ in their tarval. */
2971 h = HASH_PTR(node->attr.con.tv);
2972 h = 9*h + HASH_PTR(get_irn_mode(node));
2973 } else if (node->op == op_SymConst) {
2974 /* special value for const, as they only differ in their symbol. */
2975 h = HASH_PTR(node->attr.i.sym.type_p);
2976 h = 9*h + HASH_PTR(get_irn_mode(node));
2979 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
2980 h = irn_arity = get_irn_intra_arity(node);
2982 /* consider all in nodes... except the block if not a control flow. */
2983 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
2984 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
2988 h = 9*h + HASH_PTR(get_irn_mode(node));
2990 h = 9*h + HASH_PTR(get_irn_op(node));
2997 new_identities(void) {
2998 return new_pset(vt_cmp, N_IR_NODES);
3002 del_identities(pset *value_table) {
3003 del_pset(value_table);
3007 * Return the canonical node computing the same value as n.
3008 * Looks up the node in a hash table.
3010 * For Const nodes this is performed in the constructor, too. Const
3011 * nodes are extremely time critical because of their frequent use in
3012 * constant string arrays.
3014 static INLINE ir_node *
3015 identify (pset *value_table, ir_node *n)
3019 if (!value_table) return n;
3021 if (get_opt_reassociation()) {
3022 if (is_op_commutative(get_irn_op(n))) {
3023 ir_node *l = get_binop_left(n);
3024 ir_node *r = get_binop_right(n);
3026 /* for commutative operators perform a OP b == b OP a */
3028 set_binop_left(n, r);
3029 set_binop_right(n, l);
3034 o = pset_find (value_table, n, ir_node_hash (n));
3043 * During construction we set the op_pin_state_pinned flag in the graph right when the
3044 * optimization is performed. The flag turning on procedure global cse could
3045 * be changed between two allocations. This way we are safe.
3047 static INLINE ir_node *
3048 identify_cons (pset *value_table, ir_node *n) {
3051 n = identify(value_table, n);
3052 if (get_irn_n(old, -1) != get_irn_n(n, -1))
3053 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3058 * Return the canonical node computing the same value as n.
3059 * Looks up the node in a hash table, enters it in the table
3060 * if it isn't there yet.
3063 identify_remember (pset *value_table, ir_node *n)
3067 if (!value_table) return n;
3069 if (get_opt_reassociation()) {
3070 if (is_op_commutative(get_irn_op(n))) {
3071 ir_node *l = get_binop_left(n);
3072 ir_node *r = get_binop_right(n);
3074 /* for commutative operators perform a OP b == b OP a */
3076 set_binop_left(n, r);
3077 set_binop_right(n, l);
3082 /* lookup or insert in hash table with given hash key. */
3083 o = pset_insert (value_table, n, ir_node_hash (n));
3093 add_identities (pset *value_table, ir_node *node) {
3094 if (get_opt_cse() && (get_irn_opcode(node) != iro_Block))
3095 identify_remember (value_table, node);
3099 * garbage in, garbage out. If a node has a dead input, i.e., the
3100 * Bad node is input to the node, return the Bad node.
3102 static INLINE ir_node *
3103 gigo (ir_node *node)
3106 ir_op *op = get_irn_op(node);
3108 /* remove garbage blocks by looking at control flow that leaves the block
3109 and replacing the control flow by Bad. */
3110 if (get_irn_mode(node) == mode_X) {
3111 ir_node *block = get_nodes_block(skip_Proj(node));
3113 /* Don't optimize nodes in immature blocks. */
3114 if (!get_Block_matured(block)) return node;
3115 /* Don't optimize End, may have Bads. */
3116 if (op == op_End) return node;
3118 if (is_Block(block)) {
3119 irn_arity = get_irn_arity(block);
3120 for (i = 0; i < irn_arity; i++) {
3121 if (!is_Bad(get_irn_n(block, i)))
3124 if (i == irn_arity) return new_Bad();
3128 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
3129 blocks predecessors is dead. */
3130 if ( op != op_Block && op != op_Phi && op != op_Tuple) {
3131 irn_arity = get_irn_arity(node);
3134 * Beware: we can only read the block of a non-floating node.
3136 if (is_irn_pinned_in_irg(node) &&
3137 is_Block_dead(get_nodes_block(node)))
3140 for (i = 0; i < irn_arity; i++) {
3141 ir_node *pred = get_irn_n(node, i);
3145 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
3146 return new_Unknown(get_irn_mode(node));
3150 /* With this code we violate the agreement that local_optimize
3151 only leaves Bads in Block, Phi and Tuple nodes. */
3152 /* If Block has only Bads as predecessors it's garbage. */
3153 /* If Phi has only Bads as predecessors it's garbage. */
3154 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
3155 irn_arity = get_irn_arity(node);
3156 for (i = 0; i < irn_arity; i++) {
3157 if (!is_Bad(get_irn_n(node, i))) break;
3159 if (i == irn_arity) node = new_Bad();
3167 * These optimizations deallocate nodes from the obstack.
3168 * It can only be called if it is guaranteed that no other nodes
3169 * reference this one, i.e., right after construction of a node.
3171 * current_ir_graph must be set to the graph of the node!
3174 optimize_node(ir_node *n)
3178 opcode iro = get_irn_opcode(n);
3180 /* Always optimize Phi nodes: part of the construction. */
3181 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
3183 /* constant expression evaluation / constant folding */
3184 if (get_opt_constant_folding()) {
3185 /* neither constants nor Tuple values can be evaluated */
3186 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
3187 /* try to evaluate */
3188 tv = computed_value(n);
3189 if (tv != tarval_bad) {
3191 type *old_tp = get_irn_type(n);
3192 int i, arity = get_irn_arity(n);
3196 * Try to recover the type of the new expression.
3198 for (i = 0; i < arity && !old_tp; ++i)
3199 old_tp = get_irn_type(get_irn_n(n, i));
3202 * we MUST copy the node here temporary, because it's still needed
3203 * for DBG_OPT_CSTEVAL
3205 node_size = offsetof(ir_node, attr) + n->op->attr_size;
3206 oldn = alloca(node_size);
3208 memcpy(oldn, n, node_size);
3209 CLONE_ARR_A(ir_node *, oldn->in, n->in);
3211 /* ARG, copy the in array, we need it for statistics */
3212 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
3214 /* note the inplace edges module */
3215 edges_node_deleted(n, current_ir_graph);
3217 /* evaluation was successful -- replace the node. */
3218 obstack_free(current_ir_graph->obst, n);
3219 nw = new_Const(get_tarval_mode (tv), tv);
3221 if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv))
3222 set_Const_type(nw, old_tp);
3223 DBG_OPT_CSTEVAL(oldn, nw);
3229 /* remove unnecessary nodes */
3230 if (get_opt_constant_folding() ||
3231 (iro == iro_Phi) || /* always optimize these nodes. */
3233 (iro == iro_Proj) ||
3234 (iro == iro_Block) ) /* Flags tested local. */
3235 n = equivalent_node (n);
3237 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
3239 /* Common Subexpression Elimination.
3241 * Checks whether n is already available.
3242 * The block input is used to distinguish different subexpressions. Right
3243 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
3244 * subexpressions within a block.
3247 n = identify_cons (current_ir_graph->value_table, n);
3250 edges_node_deleted(oldn, current_ir_graph);
3252 /* We found an existing, better node, so we can deallocate the old node. */
3253 obstack_free (current_ir_graph->obst, oldn);
3258 /* Some more constant expression evaluation that does not allow to
3260 iro = get_irn_opcode(n);
3261 if (get_opt_constant_folding() ||
3262 (iro == iro_Cond) ||
3263 (iro == iro_Proj) ||
3264 (iro == iro_Sel)) /* Flags tested local. */
3265 n = transform_node (n);
3267 /* Remove nodes with dead (Bad) input.
3268 Run always for transformation induced Bads. */
3271 /* Now we have a legal, useful node. Enter it in hash table for CSE */
3272 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
3273 n = identify_remember (current_ir_graph->value_table, n);
3281 * These optimizations never deallocate nodes (in place). This can cause dead
3282 * nodes lying on the obstack. Remove these by a dead node elimination,
3283 * i.e., a copying garbage collection.
3286 optimize_in_place_2 (ir_node *n)
3290 opcode iro = get_irn_opcode(n);
3292 if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n;
3294 /* constant expression evaluation / constant folding */
3295 if (get_opt_constant_folding()) {
3296 /* neither constants nor Tuple values can be evaluated */
3297 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
3298 /* try to evaluate */
3299 tv = computed_value(n);
3300 if (tv != tarval_bad) {
3301 /* evaluation was successful -- replace the node. */
3302 type *old_tp = get_irn_type(n);
3303 int i, arity = get_irn_arity(n);
3306 * Try to recover the type of the new expression.
3308 for (i = 0; i < arity && !old_tp; ++i)
3309 old_tp = get_irn_type(get_irn_n(n, i));
3311 n = new_Const(get_tarval_mode(tv), tv);
3313 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
3314 set_Const_type(n, old_tp);
3316 DBG_OPT_CSTEVAL(oldn, n);
3322 /* remove unnecessary nodes */
3323 if (get_opt_constant_folding() ||
3324 (iro == iro_Phi) || /* always optimize these nodes. */
3325 (iro == iro_Id) || /* ... */
3326 (iro == iro_Proj) || /* ... */
3327 (iro == iro_Block) ) /* Flags tested local. */
3328 n = equivalent_node(n);
3330 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
3332 /** common subexpression elimination **/
3333 /* Checks whether n is already available. */
3334 /* The block input is used to distinguish different subexpressions. Right
3335 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
3336 subexpressions within a block. */
3337 if (get_opt_cse()) {
3338 n = identify(current_ir_graph->value_table, n);
3341 /* Some more constant expression evaluation. */
3342 iro = get_irn_opcode(n);
3343 if (get_opt_constant_folding() ||
3344 (iro == iro_Cond) ||
3345 (iro == iro_Proj) ||
3346 (iro == iro_Sel)) /* Flags tested local. */
3347 n = transform_node(n);
3349 /* Remove nodes with dead (Bad) input.
3350 Run always for transformation induced Bads. */
3353 /* Now we can verify the node, as it has no dead inputs any more. */
3356 /* Now we have a legal, useful node. Enter it in hash table for cse.
3357 Blocks should be unique anyways. (Except the successor of start:
3358 is cse with the start block!) */
3359 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
3360 n = identify_remember(current_ir_graph->value_table, n);
3366 * Wrapper for external use, set proper status bits after optimization.
3369 optimize_in_place (ir_node *n)
3371 /* Handle graph state */
3372 assert(get_irg_phase_state(current_ir_graph) != phase_building);
3374 if (get_opt_global_cse())
3375 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3376 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
3377 set_irg_outs_inconsistent(current_ir_graph);
3379 /* Maybe we could also test whether optimizing the node can
3380 change the control graph. */
3381 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
3382 set_irg_dom_inconsistent(current_ir_graph);
3383 return optimize_in_place_2 (n);
3387 * set the default ir op operations
3389 ir_op *firm_set_default_operations(ir_op *op)
3391 op = firm_set_default_computed_value(op);
3392 op = firm_set_default_equivalent_node(op);
3393 op = firm_set_default_transform_node(op);
3394 op = firm_set_default_node_cmp_attr(op);
3395 op = firm_set_default_get_type(op);