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.
28 #include "irgraph_t.h"
29 #include "iredges_t.h"
36 #include "dbginfo_t.h"
37 #include "iropt_dbg.h"
43 #include "opt_polymorphy.h"
44 #include "opt_confirms.h"
47 /* Make types visible to allow most efficient access */
48 # include "entity_t.h"
51 * return the value of a Constant
53 static tarval *computed_value_Const(ir_node *n)
55 return get_Const_tarval(n);
59 * return the value of a 'sizeof' SymConst
61 static tarval *computed_value_SymConst(ir_node *n)
65 switch (get_SymConst_kind(n)) {
66 case symconst_type_size:
67 type = get_SymConst_type(n);
68 if (get_type_state(type) == layout_fixed)
69 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
71 case symconst_type_align:
72 type = get_SymConst_type(n);
73 if (get_type_state(type) == layout_fixed)
74 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
83 * return the value of an Add
85 static tarval *computed_value_Add(ir_node *n)
87 ir_node *a = get_Add_left(n);
88 ir_node *b = get_Add_right(n);
90 tarval *ta = value_of(a);
91 tarval *tb = value_of(b);
93 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b)))
94 return tarval_add(ta, tb);
100 * return the value of a Sub
101 * Special case: a - a
103 static tarval *computed_value_Sub(ir_node *n)
105 ir_node *a = get_Sub_left(n);
106 ir_node *b = get_Sub_right(n);
111 if (a == b && !is_Bad(a))
112 return get_mode_null(get_irn_mode(n));
117 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b)))
118 return tarval_sub(ta, tb);
124 * return the value of a Carry
125 * Special : a op 0, 0 op b
127 static tarval *computed_value_Carry(ir_node *n)
129 ir_node *a = get_binop_left(n);
130 ir_node *b = get_binop_right(n);
131 ir_mode *m = get_irn_mode(n);
133 tarval *ta = value_of(a);
134 tarval *tb = value_of(b);
136 if ((ta != tarval_bad) && (tb != tarval_bad)) {
138 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
140 if ( (classify_tarval(ta) == TV_CLASSIFY_NULL)
141 || (classify_tarval(tb) == TV_CLASSIFY_NULL))
142 return get_mode_null(m);
148 * return the value of a Borrow
151 static tarval *computed_value_Borrow(ir_node *n)
153 ir_node *a = get_binop_left(n);
154 ir_node *b = get_binop_right(n);
155 ir_mode *m = get_irn_mode(n);
157 tarval *ta = value_of(a);
158 tarval *tb = value_of(b);
160 if ((ta != tarval_bad) && (tb != tarval_bad)) {
161 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
162 } else if (classify_tarval(ta) == TV_CLASSIFY_NULL) {
163 return get_mode_null(m);
169 * return the value of an unary Minus
171 static tarval *computed_value_Minus(ir_node *n)
173 ir_node *a = get_Minus_op(n);
174 tarval *ta = value_of(a);
176 if ((ta != tarval_bad) && mode_is_signed(get_irn_mode(a)))
177 return tarval_neg(ta);
183 * return the value of a Mul
185 static tarval *computed_value_Mul(ir_node *n)
187 ir_node *a = get_Mul_left(n);
188 ir_node *b = get_Mul_right(n);
190 tarval *ta = value_of(a);
191 tarval *tb = value_of(b);
193 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
194 return tarval_mul(ta, tb);
196 /* a*0 = 0 or 0*b = 0:
197 calls computed_value recursive and returns the 0 with proper
199 if ((ta != tarval_bad) && (ta == get_mode_null(get_tarval_mode(ta))))
201 if ((tb != tarval_bad) && (tb == get_mode_null(get_tarval_mode(tb))))
208 * return the value of a floating point Quot
210 static tarval *computed_value_Quot(ir_node *n)
212 ir_node *a = get_Quot_left(n);
213 ir_node *b = get_Quot_right(n);
215 tarval *ta = value_of(a);
216 tarval *tb = value_of(b);
218 /* This was missing in original implementation. Why? */
219 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
220 if (tb != get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */
221 return tarval_quo(ta, tb);
227 * calculate the value of an integer Div of two nodes
228 * Special case: 0 / b
230 static tarval *do_computed_value_Div(ir_node *a, ir_node *b)
232 tarval *ta = value_of(a);
233 tarval *tb = value_of(b);
235 /* Compute c1 / c2 or 0 / a, a != 0 */
236 if (ta != tarval_bad) {
237 if ((tb != tarval_bad) && (tb != get_mode_null(get_irn_mode(b)))) /* div by zero: return tarval_bad */
238 return tarval_div(ta, tb);
239 else if (ta == get_mode_null(get_tarval_mode(ta))) /* 0 / b == 0 */
246 * return the value of an integer Div
248 static tarval *computed_value_Div(ir_node *n)
250 return do_computed_value_Div(get_Div_left(n), get_Div_right(n));
254 * calculate the value of an integer Mod of two nodes
255 * Special case: a % 1
257 static tarval *do_computed_value_Mod(ir_node *a, ir_node *b)
259 tarval *ta = value_of(a);
260 tarval *tb = value_of(b);
262 /* Compute c1 % c2 or a % 1 */
263 if (tb != tarval_bad) {
264 if ((ta != tarval_bad) && (tb != get_mode_null(get_tarval_mode(tb)))) /* div by zero: return tarval_bad */
265 return tarval_mod(ta, tb);
266 else if (tb == get_mode_one(get_tarval_mode(tb))) /* x mod 1 == 0 */
267 return get_mode_null(get_irn_mode(a));
274 * return the value of an integer Mod
276 static tarval *computed_value_Mod(ir_node *n)
278 return do_computed_value_Mod(get_Mod_left(n), get_Mod_right(n));
282 * return the value of an Abs
284 static tarval *computed_value_Abs(ir_node *n)
286 ir_node *a = get_Abs_op(n);
287 tarval *ta = value_of(a);
289 if (ta != tarval_bad)
290 return tarval_abs(ta);
296 * return the value of an And
297 * Special case: a & 0, 0 & b
299 static tarval *computed_value_And(ir_node *n)
301 ir_node *a = get_And_left(n);
302 ir_node *b = get_And_right(n);
304 tarval *ta = value_of(a);
305 tarval *tb = value_of(b);
307 if ((ta != tarval_bad) && (tb != tarval_bad)) {
308 return tarval_and (ta, tb);
312 if ( (classify_tarval ((v = ta)) == TV_CLASSIFY_NULL)
313 || (classify_tarval ((v = tb)) == TV_CLASSIFY_NULL)) {
321 * return the value of an Or
322 * Special case: a | 1...1, 1...1 | b
324 static tarval *computed_value_Or(ir_node *n)
326 ir_node *a = get_Or_left(n);
327 ir_node *b = get_Or_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_or (ta, tb);
336 if ( (classify_tarval ((v = ta)) == TV_CLASSIFY_ALL_ONE)
337 || (classify_tarval ((v = tb)) == TV_CLASSIFY_ALL_ONE)) {
345 * return the value of an Eor
347 static tarval *computed_value_Eor(ir_node *n)
349 ir_node *a = get_Eor_left(n);
350 ir_node *b = get_Eor_right(n);
355 return get_mode_null(get_irn_mode(n));
360 if ((ta != tarval_bad) && (tb != tarval_bad)) {
361 return tarval_eor (ta, tb);
367 * return the value of a Not
369 static tarval *computed_value_Not(ir_node *n)
371 ir_node *a = get_Not_op(n);
372 tarval *ta = value_of(a);
374 if (ta != tarval_bad)
375 return tarval_not(ta);
381 * return the value of a Shl
383 static tarval *computed_value_Shl(ir_node *n)
385 ir_node *a = get_Shl_left(n);
386 ir_node *b = get_Shl_right(n);
388 tarval *ta = value_of(a);
389 tarval *tb = value_of(b);
391 if ((ta != tarval_bad) && (tb != tarval_bad)) {
392 return tarval_shl (ta, tb);
398 * return the value of a Shr
400 static tarval *computed_value_Shr(ir_node *n)
402 ir_node *a = get_Shr_left(n);
403 ir_node *b = get_Shr_right(n);
405 tarval *ta = value_of(a);
406 tarval *tb = value_of(b);
408 if ((ta != tarval_bad) && (tb != tarval_bad)) {
409 return tarval_shr (ta, tb);
415 * return the value of a Shrs
417 static tarval *computed_value_Shrs(ir_node *n)
419 ir_node *a = get_Shrs_left(n);
420 ir_node *b = get_Shrs_right(n);
422 tarval *ta = value_of(a);
423 tarval *tb = value_of(b);
425 if ((ta != tarval_bad) && (tb != tarval_bad)) {
426 return tarval_shrs (ta, tb);
432 * return the value of a Rot
434 static tarval *computed_value_Rot(ir_node *n)
436 ir_node *a = get_Rot_left(n);
437 ir_node *b = get_Rot_right(n);
439 tarval *ta = value_of(a);
440 tarval *tb = value_of(b);
442 if ((ta != tarval_bad) && (tb != tarval_bad)) {
443 return tarval_rot (ta, tb);
449 * return the value of a Conv
451 static tarval *computed_value_Conv(ir_node *n)
453 ir_node *a = get_Conv_op(n);
454 tarval *ta = value_of(a);
456 if (ta != tarval_bad)
457 return tarval_convert_to(ta, get_irn_mode(n));
463 * return the value of a Proj(Cmp)
465 * This performs a first step of unreachable code elimination.
466 * Proj can not be computed, but folding a Cmp above the Proj here is
467 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
469 * There are several case where we can evaluate a Cmp node, see later.
471 static tarval *computed_value_Proj_Cmp(ir_node *n)
473 ir_node *a = get_Proj_pred(n);
474 ir_node *aa = get_Cmp_left(a);
475 ir_node *ab = get_Cmp_right(a);
476 long proj_nr = get_Proj_proj(n);
479 * BEWARE: a == a is NOT always True for floating Point values, as
480 * NaN != NaN is defined, so we must check this here.
483 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
486 /* This is a trick with the bits used for encoding the Cmp
487 Proj numbers, the following statement is not the same:
488 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
489 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
492 tarval *taa = value_of(aa);
493 tarval *tab = value_of(ab);
494 ir_mode *mode = get_irn_mode(aa);
497 * The predecessors of Cmp are target values. We can evaluate
500 if ((taa != tarval_bad) && (tab != tarval_bad)) {
501 /* strange checks... */
502 pn_Cmp flags = tarval_cmp(taa, tab);
503 if (flags != pn_Cmp_False) {
504 return new_tarval_from_long (proj_nr & flags, mode_b);
507 /* for integer values, we can check against MIN/MAX */
508 else if (mode_is_int(mode)) {
509 /* MIN <=/> x. This results in true/false. */
510 if (taa == get_mode_min(mode)) {
511 /* a compare with the MIN value */
512 if (proj_nr == pn_Cmp_Le)
513 return get_tarval_b_true();
514 else if (proj_nr == pn_Cmp_Gt)
515 return get_tarval_b_false();
517 /* x >=/< MIN. This results in true/false. */
519 if (tab == get_mode_min(mode)) {
520 /* a compare with the MIN value */
521 if (proj_nr == pn_Cmp_Ge)
522 return get_tarval_b_true();
523 else if (proj_nr == pn_Cmp_Lt)
524 return get_tarval_b_false();
526 /* MAX >=/< x. This results in true/false. */
527 else if (taa == get_mode_max(mode)) {
528 if (proj_nr == pn_Cmp_Ge)
529 return get_tarval_b_true();
530 else if (proj_nr == pn_Cmp_Lt)
531 return get_tarval_b_false();
533 /* x <=/> MAX. This results in true/false. */
534 else if (tab == get_mode_max(mode)) {
535 if (proj_nr == pn_Cmp_Le)
536 return get_tarval_b_true();
537 else if (proj_nr == pn_Cmp_Gt)
538 return get_tarval_b_false();
542 * The predecessors are Allocs or (void*)(0) constants. Allocs never
543 * return NULL, they raise an exception. Therefore we can predict
547 ir_node *aaa = skip_Id(skip_Proj(aa));
548 ir_node *aba = skip_Id(skip_Proj(ab));
550 if ( ( (/* aa is ProjP and aaa is Alloc */
551 (get_irn_op(aa) == op_Proj)
552 && (mode_is_reference(get_irn_mode(aa)))
553 && (get_irn_op(aaa) == op_Alloc))
554 && ( (/* ab is NULL */
555 (get_irn_op(ab) == op_Const)
556 && (mode_is_reference(get_irn_mode(ab)))
557 && (get_Const_tarval(ab) == get_mode_null(get_irn_mode(ab))))
558 || (/* ab is other Alloc */
559 (get_irn_op(ab) == op_Proj)
560 && (mode_is_reference(get_irn_mode(ab)))
561 && (get_irn_op(aba) == op_Alloc)
563 || (/* aa is NULL and aba is Alloc */
564 (get_irn_op(aa) == op_Const)
565 && (mode_is_reference(get_irn_mode(aa)))
566 && (get_Const_tarval(aa) == get_mode_null(get_irn_mode(aa)))
567 && (get_irn_op(ab) == op_Proj)
568 && (mode_is_reference(get_irn_mode(ab)))
569 && (get_irn_op(aba) == op_Alloc)))
571 return new_tarval_from_long(proj_nr & pn_Cmp_Ne, mode_b);
574 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
578 * return the value of a Proj, handle Proj(Cmp), Proj(Div), Proj(Mod), Proj(DivMod)
580 static tarval *computed_value_Proj(ir_node *n)
582 ir_node *a = get_Proj_pred(n);
585 switch (get_irn_opcode(a)) {
587 return computed_value_Proj_Cmp(n);
590 /* compute either the Div or the Mod part */
591 proj_nr = get_Proj_proj(n);
592 if (proj_nr == pn_DivMod_res_div)
593 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
594 else if (proj_nr == pn_DivMod_res_mod)
595 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
599 if (get_Proj_proj(n) == pn_Div_res)
600 return computed_value(a);
604 if (get_Proj_proj(n) == pn_Mod_res)
605 return computed_value(a);
615 * calculate the value of a Mux: can be evaluated, if the
616 * sel and the right input are known
618 static tarval *computed_value_Mux(ir_node *n)
620 ir_node *sel = get_Mux_sel(n);
621 tarval *ts = value_of(sel);
623 if (ts == get_tarval_b_true()) {
624 ir_node *v = get_Mux_true(n);
627 else if (ts == get_tarval_b_false()) {
628 ir_node *v = get_Mux_false(n);
635 * Calculate the value of a Psi: can be evaluated, if a condition is true
636 * and all previous conditions are false. If all conditions are false
637 * we evaluate to the default one.
639 static tarval *computed_value_Psi(ir_node *n)
642 return computed_value_Mux(n);
647 * calculate the value of a Confirm: can be evaluated,
648 * if it has the form Confirm(x, '=', Const).
650 static tarval *computed_value_Confirm(ir_node *n)
652 return get_Confirm_cmp(n) == pn_Cmp_Eq ?
653 value_of(get_Confirm_bound(n)) : tarval_bad;
657 * If the parameter n can be computed, return its value, else tarval_bad.
658 * Performs constant folding.
660 * @param n The node this should be evaluated
662 tarval *computed_value(ir_node *n)
664 if (n->op->ops.computed_value)
665 return n->op->ops.computed_value(n);
670 * set the default computed_value evaluator in an ir_op_ops.
672 * @param code the opcode for the default operation
673 * @param ops the operations initialized
678 static ir_op_ops *firm_set_default_computed_value(opcode code, ir_op_ops *ops)
682 ops->computed_value = computed_value_##a; \
720 * Returns a equivalent block for another block.
721 * If the block has only one predecessor, this is
722 * the equivalent one. If the only predecessor of a block is
723 * the block itself, this is a dead block.
725 * If both predecessors of a block are the branches of a binary
726 * Cond, the equivalent block is Cond's block.
728 * If all predecessors of a block are bad or lies in a dead
729 * block, the current block is dead as well.
731 * Note, that blocks are NEVER turned into Bad's, instead
732 * the dead_block flag is set. So, never test for is_Bad(block),
733 * always use is_dead_Block(block).
735 static ir_node *equivalent_node_Block(ir_node *n)
738 int n_preds = get_Block_n_cfgpreds(n);
740 /* The Block constructor does not call optimize, but mature_immBlock
741 calls the optimization. */
742 assert(get_Block_matured(n));
744 /* Straightening: a single entry Block following a single exit Block
745 can be merged, if it is not the Start block. */
746 /* !!! Beware, all Phi-nodes of n must have been optimized away.
747 This should be true, as the block is matured before optimize is called.
748 But what about Phi-cycles with the Phi0/Id that could not be resolved?
749 Remaining Phi nodes are just Ids. */
750 if ((n_preds == 1) && (get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp)) {
751 ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0));
752 if (predblock == oldn) {
753 /* Jmp jumps into the block it is in -- deal self cycle. */
754 n = set_Block_dead(n);
755 DBG_OPT_DEAD_BLOCK(oldn, n);
756 } else if (get_opt_control_flow_straightening()) {
758 DBG_OPT_STG(oldn, n);
761 else if ((n_preds == 1) &&
762 (get_irn_op(skip_Proj(get_Block_cfgpred(n, 0))) == op_Cond)) {
763 ir_node *predblock = get_Block_cfgpred_block(n, 0);
764 if (predblock == oldn) {
765 /* Jmp jumps into the block it is in -- deal self cycle. */
766 n = set_Block_dead(n);
767 DBG_OPT_DEAD_BLOCK(oldn, n);
770 else if ((n_preds == 2) &&
771 (get_opt_control_flow_weak_simplification())) {
772 /* Test whether Cond jumps twice to this block
773 * The more general case which more than 2 predecessors is handles
774 * in optimize_cf(), we handle only this special case for speed here.
776 ir_node *a = get_Block_cfgpred(n, 0);
777 ir_node *b = get_Block_cfgpred(n, 1);
779 if ((get_irn_op(a) == op_Proj) &&
780 (get_irn_op(b) == op_Proj) &&
781 (get_Proj_pred(a) == get_Proj_pred(b)) &&
782 (get_irn_op(get_Proj_pred(a)) == op_Cond) &&
783 (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) {
784 /* Also a single entry Block following a single exit Block. Phis have
785 twice the same operand and will be optimized away. */
786 n = get_nodes_block(get_Proj_pred(a));
787 DBG_OPT_IFSIM1(oldn, a, b, n);
790 else if (get_opt_unreachable_code() &&
791 (n != get_irg_start_block(current_ir_graph)) &&
792 (n != get_irg_end_block(current_ir_graph)) ) {
795 /* If all inputs are dead, this block is dead too, except if it is
796 the start or end block. This is one step of unreachable code
798 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
799 ir_node *pred = get_Block_cfgpred(n, i);
802 if (is_Bad(pred)) continue;
803 pred_blk = get_nodes_block(skip_Proj(pred));
805 if (is_Block_dead(pred_blk)) continue;
808 /* really found a living input */
813 n = set_Block_dead(n);
814 DBG_OPT_DEAD_BLOCK(oldn, n);
822 * Returns a equivalent node for a Jmp, a Bad :-)
823 * Of course this only happens if the Block of the Jmp is dead.
825 static ir_node *equivalent_node_Jmp(ir_node *n)
827 /* unreachable code elimination */
828 if (is_Block_dead(get_nodes_block(n)))
834 /** Raise is handled in the same way as Jmp. */
835 #define equivalent_node_Raise equivalent_node_Jmp
838 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
839 See transform_node_Proj_Cond(). */
842 * optimize operations that are commutative and have neutral 0,
843 * so a op 0 = 0 op a = a.
845 static ir_node *equivalent_node_neutral_zero(ir_node *n)
849 ir_node *a = get_binop_left(n);
850 ir_node *b = get_binop_right(n);
855 /* After running compute_node there is only one constant predecessor.
856 Find this predecessors value and remember the other node: */
857 if ((tv = value_of(a)) != tarval_bad) {
859 } else if ((tv = value_of(b)) != tarval_bad) {
864 /* If this predecessors constant value is zero, the operation is
865 * unnecessary. Remove it.
867 * Beware: If n is a Add, the mode of on and n might be different
868 * which happens in this rare construction: NULL + 3.
869 * Then, a Conv would be needed which we cannot include here.
871 if (classify_tarval (tv) == TV_CLASSIFY_NULL) {
872 if (get_irn_mode(on) == get_irn_mode(n)) {
875 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
883 * Eor is commutative and has neutral 0.
885 #define equivalent_node_Eor equivalent_node_neutral_zero
888 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
890 * The second one looks strange, but this construct
891 * is used heavily in the LCC sources :-).
893 * Beware: The Mode of an Add may be different than the mode of its
894 * predecessors, so we could not return a predecessors in all cases.
896 static ir_node *equivalent_node_Add(ir_node *n)
899 ir_node *left, *right;
901 n = equivalent_node_neutral_zero(n);
905 left = get_Add_left(n);
906 right = get_Add_right(n);
908 if (get_irn_op(left) == op_Sub) {
909 if (get_Sub_right(left) == right) {
912 n = get_Sub_left(left);
913 if (get_irn_mode(oldn) == get_irn_mode(n)) {
914 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
919 if (get_irn_op(right) == op_Sub) {
920 if (get_Sub_right(right) == left) {
923 n = get_Sub_left(right);
924 if (get_irn_mode(oldn) == get_irn_mode(n)) {
925 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
934 * optimize operations that are not commutative but have neutral 0 on left,
937 static ir_node *equivalent_node_left_zero(ir_node *n)
941 ir_node *a = get_binop_left(n);
942 ir_node *b = get_binop_right(n);
944 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
947 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
953 #define equivalent_node_Shl equivalent_node_left_zero
954 #define equivalent_node_Shr equivalent_node_left_zero
955 #define equivalent_node_Shrs equivalent_node_left_zero
956 #define equivalent_node_Rot equivalent_node_left_zero
959 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
961 * The second one looks strange, but this construct
962 * is used heavily in the LCC sources :-).
964 * Beware: The Mode of a Sub may be different than the mode of its
965 * predecessors, so we could not return a predecessors in all cases.
967 static ir_node *equivalent_node_Sub(ir_node *n)
971 ir_node *a = get_Sub_left(n);
972 ir_node *b = get_Sub_right(n);
974 /* Beware: modes might be different */
975 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
976 if (get_irn_mode(n) == get_irn_mode(a)) {
979 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
982 else if (get_irn_op(a) == op_Add) {
983 ir_mode *mode = get_irn_mode(n);
985 if (mode_wrap_around(mode)) {
986 ir_node *left = get_Add_left(a);
987 ir_node *right = get_Add_right(a);
990 if (get_irn_mode(n) == get_irn_mode(right)) {
992 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
995 else if (right == b) {
996 if (get_irn_mode(n) == get_irn_mode(left)) {
998 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
1009 * Optimize an "idempotent unary op", ie op(op(n)) = n.
1012 * -(-a) == a, but might overflow two times.
1013 * We handle it anyway here but the better way would be a
1014 * flag. This would be needed for Pascal for instance.
1016 static ir_node *equivalent_node_idempotent_unop(ir_node *n)
1019 ir_node *pred = get_unop_op(n);
1021 /* optimize symmetric unop */
1022 if (get_irn_op(pred) == get_irn_op(n)) {
1023 n = get_unop_op(pred);
1024 DBG_OPT_ALGSIM2(oldn, pred, n);
1029 /** Not(Not(x)) == x */
1030 #define equivalent_node_Not equivalent_node_idempotent_unop
1032 /** --x == x ??? Is this possible or can --x raise an
1033 out of bounds exception if min =! max? */
1034 #define equivalent_node_Minus equivalent_node_idempotent_unop
1037 * Optimize a * 1 = 1 * a = a.
1039 static ir_node *equivalent_node_Mul(ir_node *n)
1043 ir_node *a = get_Mul_left(n);
1044 ir_node *b = get_Mul_right(n);
1046 /* Mul is commutative and has again an other neutral element. */
1047 if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) {
1049 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1050 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) {
1052 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1058 * Optimize a / 1 = a.
1060 static ir_node *equivalent_node_Div(ir_node *n)
1062 ir_node *a = get_Div_left(n);
1063 ir_node *b = get_Div_right(n);
1065 /* Div is not commutative. */
1066 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1067 /* Turn Div into a tuple (mem, bad, a) */
1068 ir_node *mem = get_Div_mem(n);
1069 turn_into_tuple(n, pn_Div_max);
1070 set_Tuple_pred(n, pn_Div_M, mem);
1071 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
1072 set_Tuple_pred(n, pn_Div_res, a);
1078 * Optimize a / 1 = a.
1080 static ir_node *equivalent_node_DivMod(ir_node *n)
1082 ir_node *a = get_DivMod_left(n);
1083 ir_node *b = get_DivMod_right(n);
1085 /* Div is not commutative. */
1086 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1087 /* Turn DivMod into a tuple (mem, bad, a, 0) */
1088 ir_node *mem = get_Div_mem(n);
1089 ir_mode *mode = get_irn_mode(b);
1091 turn_into_tuple(n, pn_DivMod_max);
1092 set_Tuple_pred(n, pn_DivMod_M, mem);
1093 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1094 set_Tuple_pred(n, pn_DivMod_res_div, a);
1095 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1101 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1103 static ir_node *equivalent_node_Or(ir_node *n)
1107 ir_node *a = get_Or_left(n);
1108 ir_node *b = get_Or_right(n);
1111 n = a; /* Or has it's own neutral element */
1112 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1113 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) {
1115 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1116 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
1118 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1125 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a.
1127 static ir_node *equivalent_node_And(ir_node *n)
1131 ir_node *a = get_And_left(n);
1132 ir_node *b = get_And_right(n);
1135 n = a; /* And has it's own neutral element */
1136 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1137 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) {
1139 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1140 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) {
1142 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1148 * Try to remove useless Conv's:
1150 static ir_node *equivalent_node_Conv(ir_node *n)
1153 ir_node *a = get_Conv_op(n);
1156 ir_mode *n_mode = get_irn_mode(n);
1157 ir_mode *a_mode = get_irn_mode(a);
1159 if (n_mode == a_mode) { /* No Conv necessary */
1161 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1162 } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */
1166 n_mode = get_irn_mode(n);
1167 b_mode = get_irn_mode(b);
1169 if (n_mode == b_mode) {
1170 if (n_mode == mode_b) {
1171 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1172 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1174 else if (mode_is_int(n_mode) || mode_is_character(n_mode)) {
1175 if (smaller_mode(b_mode, a_mode)){
1176 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1177 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1186 * A Cast may be removed if the type of the previous node
1187 * is already the type of the Cast.
1189 static ir_node *equivalent_node_Cast(ir_node *n) {
1191 ir_node *pred = get_Cast_op(n);
1193 if (get_irn_type(pred) == get_Cast_type(n)) {
1195 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1201 Several optimizations:
1202 - no Phi in start block.
1203 - remove Id operators that are inputs to Phi
1204 - fold Phi-nodes, iff they have only one predecessor except
1207 static ir_node *equivalent_node_Phi(ir_node *n)
1212 ir_node *block = NULL; /* to shutup gcc */
1213 ir_node *first_val = NULL; /* to shutup gcc */
1215 if (!get_opt_normalize()) return n;
1217 n_preds = get_Phi_n_preds(n);
1219 block = get_nodes_block(n);
1220 /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
1221 assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
1222 if ((is_Block_dead(block)) || /* Control dead */
1223 (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */
1224 return new_Bad(); /* in the Start Block. */
1226 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1228 /* If the Block has a Bad pred, we also have one. */
1229 for (i = 0; i < n_preds; ++i)
1230 if (is_Bad(get_Block_cfgpred(block, i)))
1231 set_Phi_pred(n, i, new_Bad());
1233 /* Find first non-self-referencing input */
1234 for (i = 0; i < n_preds; ++i) {
1235 first_val = get_Phi_pred(n, i);
1236 if ( (first_val != n) /* not self pointer */
1238 && (! is_Bad(first_val))
1240 ) { /* value not dead */
1241 break; /* then found first value. */
1246 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1250 /* search for rest of inputs, determine if any of these
1251 are non-self-referencing */
1252 while (++i < n_preds) {
1253 ir_node *scnd_val = get_Phi_pred(n, i);
1254 if ( (scnd_val != n)
1255 && (scnd_val != first_val)
1257 && (! is_Bad(scnd_val))
1265 /* Fold, if no multiple distinct non-self-referencing inputs */
1267 DBG_OPT_PHI(oldn, n);
1273 Several optimizations:
1274 - no Sync in start block.
1275 - fold Sync-nodes, iff they have only one predecessor except
1277 @fixme: are there loop's in Sync's
1279 static ir_node *equivalent_node_Sync(ir_node *n)
1284 ir_node *first_val = NULL; /* to shutup gcc */
1286 if (!get_opt_normalize()) return n;
1288 n_preds = get_Sync_n_preds(n);
1290 /* Find first non-self-referencing input */
1291 for (i = 0; i < n_preds; ++i) {
1292 first_val = get_Sync_pred(n, i);
1293 if ((first_val != n) /* not self pointer */ &&
1294 (! is_Bad(first_val))
1295 ) { /* value not dead */
1296 break; /* then found first value. */
1301 /* A totally Bad or self-referencing Sync (we didn't break the above loop) */
1304 /* search the rest of inputs, determine if any of these
1305 are non-self-referencing */
1306 while (++i < n_preds) {
1307 ir_node *scnd_val = get_Sync_pred(n, i);
1308 if ((scnd_val != n) &&
1309 (scnd_val != first_val) &&
1310 (! is_Bad(scnd_val))
1316 /* Fold, if no multiple distinct non-self-referencing inputs */
1318 DBG_OPT_SYNC(oldn, n);
1324 * optimize Proj(Tuple) and gigo() for ProjX in Bad block,
1325 * ProjX(Load) and ProjX(Store)
1327 static ir_node *equivalent_node_Proj(ir_node *n)
1331 ir_node *a = get_Proj_pred(n);
1333 if ( get_irn_op(a) == op_Tuple) {
1334 /* Remove the Tuple/Proj combination. */
1335 if ( get_Proj_proj(n) <= get_Tuple_n_preds(a) ) {
1336 n = get_Tuple_pred(a, get_Proj_proj(n));
1337 DBG_OPT_TUPLE(oldn, a, n);
1339 assert(0); /* This should not happen! */
1343 else if (get_irn_mode(n) == mode_X) {
1344 if (is_Block_dead(get_nodes_block(skip_Proj(n)))) {
1345 /* Remove dead control flow -- early gigo(). */
1348 else if (get_opt_ldst_only_null_ptr_exceptions()) {
1349 ir_op *op = get_irn_op(a);
1351 if (op == op_Load || op == op_Store) {
1352 /* get the load/store address */
1353 ir_node *addr = get_irn_n(a, 1);
1354 if (value_not_null(addr)) {
1355 /* this node may float if it did not depend on a Confirm */
1356 set_irn_pinned(a, op_pin_state_floats);
1370 static ir_node *equivalent_node_Id(ir_node *n)
1376 } while (get_irn_op(n) == op_Id);
1378 DBG_OPT_ID(oldn, n);
1385 static ir_node *equivalent_node_Mux(ir_node *n)
1387 ir_node *oldn = n, *sel = get_Mux_sel(n);
1388 tarval *ts = value_of(sel);
1390 /* Mux(true, f, t) == t */
1391 if (ts == tarval_b_true) {
1392 n = get_Mux_true(n);
1393 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1395 /* Mux(false, f, t) == f */
1396 else if (ts == tarval_b_false) {
1397 n = get_Mux_false(n);
1398 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1400 /* Mux(v, x, x) == x */
1401 else if (get_Mux_false(n) == get_Mux_true(n)) {
1402 n = get_Mux_true(n);
1403 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1405 else if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(get_irn_mode(n))) {
1406 ir_node *cmp = get_Proj_pred(sel);
1407 long proj_nr = get_Proj_proj(sel);
1408 ir_node *b = get_Mux_false(n);
1409 ir_node *a = get_Mux_true(n);
1412 * Note: normalization puts the constant on the right site,
1413 * so we check only one case.
1415 * Note further that these optimization work even for floating point
1416 * with NaN's because -NaN == NaN.
1417 * However, if +0 and -0 is handled differently, we cannot use the first one.
1419 if (get_irn_op(cmp) == op_Cmp && get_Cmp_left(cmp) == a) {
1420 if (classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
1421 /* Mux(a CMP 0, X, a) */
1422 if (get_irn_op(b) == op_Minus && get_Minus_op(b) == a) {
1423 /* Mux(a CMP 0, -a, a) */
1424 if (proj_nr == pn_Cmp_Eq) {
1425 /* Mux(a == 0, -a, a) ==> -a */
1427 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1429 else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1430 /* Mux(a != 0, -a, a) ==> a */
1432 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1435 else if (classify_Const(b) == CNST_NULL) {
1436 /* Mux(a CMP 0, 0, a) */
1437 if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1438 /* Mux(a != 0, 0, a) ==> a */
1440 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1442 else if (proj_nr == pn_Cmp_Eq) {
1443 /* Mux(a == 0, 0, a) ==> 0 */
1445 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1455 * Returns a equivalent node of a Psi: if a condition is true
1456 * and all previous conditions are false we know its value.
1457 * If all conditions are false its value is the default one.
1459 static ir_node *equivalent_node_Psi(ir_node *n) {
1461 return equivalent_node_Mux(n);
1466 * Optimize -a CMP -b into b CMP a.
1467 * This works only for for modes where unary Minus
1469 * Note that two-complement integers can Overflow
1470 * so it will NOT work.
1472 static ir_node *equivalent_node_Cmp(ir_node *n)
1474 ir_node *left = get_Cmp_left(n);
1475 ir_node *right = get_Cmp_right(n);
1477 if (get_irn_op(left) == op_Minus && get_irn_op(right) == op_Minus &&
1478 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1479 left = get_Minus_op(left);
1480 right = get_Minus_op(right);
1481 set_Cmp_left(n, right);
1482 set_Cmp_right(n, left);
1488 * Remove Confirm nodes if setting is on.
1489 * Replace Confirms(x, '=', Constlike) by Constlike.
1491 static ir_node *equivalent_node_Confirm(ir_node *n)
1493 ir_node *pred = get_Confirm_value(n);
1494 pn_Cmp pnc = get_Confirm_cmp(n);
1496 if (get_irn_op(pred) == op_Confirm && pnc == get_Confirm_cmp(pred)) {
1498 * rare case: two identical Confirms one after another,
1499 * replace the second one with the first.
1503 if (pnc == pn_Cmp_Eq) {
1504 ir_node *bound = get_Confirm_bound(n);
1507 * Optimize a rare case:
1508 * Confirm(x, '=', Constlike) ==> Constlike
1510 if (is_irn_constlike(bound)) {
1511 DBG_OPT_CONFIRM(n, bound);
1515 return get_opt_remove_confirm() ? get_Confirm_value(n) : n;
1519 * Optimize CopyB(mem, x, x) into a Nop
1521 static ir_node *equivalent_node_CopyB(ir_node *n)
1523 ir_node *a = get_CopyB_dst(n);
1524 ir_node *b = get_CopyB_src(n);
1527 /* Turn CopyB into a tuple (mem, bad, bad) */
1528 ir_node *mem = get_CopyB_mem(n);
1529 turn_into_tuple(n, pn_CopyB_max);
1530 set_Tuple_pred(n, pn_CopyB_M, mem);
1531 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1532 set_Tuple_pred(n, pn_CopyB_M_except, new_Bad());
1538 * Optimize Bounds(idx, idx, upper) into idx.
1540 static ir_node *equivalent_node_Bound(ir_node *n)
1542 ir_node *idx = get_Bound_index(n);
1543 ir_node *lower = get_Bound_lower(n);
1546 /* By definition lower < upper, so if idx == lower -->
1547 lower <= idx && idx < upper */
1549 /* Turn Bound into a tuple (mem, bad, idx) */
1553 ir_node *pred = skip_Proj(idx);
1555 if (get_irn_op(pred) == op_Bound) {
1557 * idx was Bounds_check previously, it is still valid if
1558 * lower <= pred_lower && pred_upper <= upper.
1560 ir_node *upper = get_Bound_upper(n);
1561 if (get_Bound_lower(pred) == lower &&
1562 get_Bound_upper(pred) == upper) {
1564 * One could expect that we simply return the previous
1565 * Bound here. However, this would be wrong, as we could
1566 * add an exception Proj to a new location than.
1567 * So, we must turn in into a tuple
1574 /* Turn Bound into a tuple (mem, bad, idx) */
1575 ir_node *mem = get_Bound_mem(n);
1576 turn_into_tuple(n, pn_Bound_max);
1577 set_Tuple_pred(n, pn_Bound_M, mem);
1578 set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */
1579 set_Tuple_pred(n, pn_Bound_res, idx);
1585 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1586 * perform no actual computation, as, e.g., the Id nodes. It does not create
1587 * new nodes. It is therefore safe to free n if the node returned is not n.
1588 * If a node returns a Tuple we can not just skip it. If the size of the
1589 * in array fits, we transform n into a tuple (e.g., Div).
1592 equivalent_node(ir_node *n)
1594 if (n->op->ops.equivalent_node)
1595 return n->op->ops.equivalent_node(n);
1600 * sets the default equivalent node operation for an ir_op_ops.
1602 * @param code the opcode for the default operation
1603 * @param ops the operations initialized
1608 static ir_op_ops *firm_set_default_equivalent_node(opcode code, ir_op_ops *ops)
1612 ops->equivalent_node = equivalent_node_##a; \
1654 * Do node specific optimizations of nodes predecessors.
1657 optimize_preds(ir_node *n) {
1658 ir_node *a = NULL, *b = NULL;
1660 /* get the operands we will work on for simple cases. */
1662 a = get_binop_left(n);
1663 b = get_binop_right(n);
1664 } else if (is_unop(n)) {
1668 switch (get_irn_opcode(n)) {
1671 /* We don't want Cast as input to Cmp. */
1672 if (get_irn_op(a) == op_Cast) {
1676 if (get_irn_op(b) == op_Cast) {
1678 set_Cmp_right(n, b);
1687 * Returns non-zero if all Phi predecessors are constants
1689 static int is_const_Phi(ir_node *phi) {
1692 for (i = get_irn_arity(phi) - 1; i >= 0; --i)
1693 if (! is_Const(get_irn_n(phi, i)))
1699 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1700 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1701 * If possible, remove the Conv's.
1703 static ir_node *transform_node_AddSub(ir_node *n)
1705 ir_mode *mode = get_irn_mode(n);
1707 if (mode_is_reference(mode)) {
1708 ir_node *left = get_binop_left(n);
1709 ir_node *right = get_binop_right(n);
1710 int ref_bits = get_mode_size_bits(mode);
1712 if (get_irn_op(left) == op_Conv) {
1713 ir_mode *mode = get_irn_mode(left);
1714 int bits = get_mode_size_bits(mode);
1716 if (ref_bits == bits &&
1717 mode_is_int(mode) &&
1718 get_mode_arithmetic(mode) == irma_twos_complement) {
1719 ir_node *pre = get_Conv_op(left);
1720 ir_mode *pre_mode = get_irn_mode(pre);
1722 if (mode_is_int(pre_mode) &&
1723 get_mode_size_bits(pre_mode) == bits &&
1724 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1725 /* ok, this conv just changes to sign, moreover the calculation
1726 * is done with same number of bits as our address mode, so
1727 * we can ignore the conv as address calculation can be viewed
1728 * as either signed or unsigned
1730 set_binop_left(n, pre);
1735 if (get_irn_op(right) == op_Conv) {
1736 ir_mode *mode = get_irn_mode(right);
1737 int bits = get_mode_size_bits(mode);
1739 if (ref_bits == bits &&
1740 mode_is_int(mode) &&
1741 get_mode_arithmetic(mode) == irma_twos_complement) {
1742 ir_node *pre = get_Conv_op(right);
1743 ir_mode *pre_mode = get_irn_mode(pre);
1745 if (mode_is_int(pre_mode) &&
1746 get_mode_size_bits(pre_mode) == bits &&
1747 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1748 /* ok, this conv just changes to sign, moreover the calculation
1749 * is done with same number of bits as our address mode, so
1750 * we can ignore the conv as address calculation can be viewed
1751 * as either signed or unsigned
1753 set_binop_right(n, pre);
1762 * Do the AddSub optimization, then Transform
1763 * Add(a,a) -> Mul(a, 2)
1764 * Add(Mul(a, x), a) -> Mul(a, x+1)
1765 * if the mode is integer or float.
1766 * Transform Add(a,-b) into Sub(a,b).
1767 * Reassociation might fold this further.
1769 static ir_node *transform_node_Add(ir_node *n)
1774 n = transform_node_AddSub(n);
1776 mode = get_irn_mode(n);
1777 if (mode_is_num(mode)) {
1778 ir_node *a = get_Add_left(n);
1779 ir_node *b = get_Add_right(n);
1782 ir_node *block = get_irn_n(n, -1);
1785 get_irn_dbg_info(n),
1789 new_r_Const_long(current_ir_graph, block, mode, 2),
1791 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
1793 else if (get_irn_op(a) == op_Minus) {
1795 get_irn_dbg_info(n),
1801 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1803 else if (get_irn_op(b) == op_Minus) {
1805 get_irn_dbg_info(n),
1811 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1813 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1814 else if (!get_opt_reassociation() && get_irn_op(a) == op_Mul) {
1815 ir_node *ma = get_Mul_left(a);
1816 ir_node *mb = get_Mul_right(a);
1819 ir_node *blk = get_irn_n(n, -1);
1821 get_irn_dbg_info(n), current_ir_graph, blk,
1824 get_irn_dbg_info(n), current_ir_graph, blk,
1826 new_r_Const_long(current_ir_graph, blk, mode, 1),
1829 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1832 ir_node *blk = get_irn_n(n, -1);
1834 get_irn_dbg_info(n), current_ir_graph, blk,
1837 get_irn_dbg_info(n), current_ir_graph, blk,
1839 new_r_Const_long(current_ir_graph, blk, mode, 1),
1842 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1845 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1846 else if (!get_opt_reassociation() && get_irn_op(b) == op_Mul) {
1847 ir_node *ma = get_Mul_left(b);
1848 ir_node *mb = get_Mul_right(b);
1851 ir_node *blk = get_irn_n(n, -1);
1853 get_irn_dbg_info(n), current_ir_graph, blk,
1856 get_irn_dbg_info(n), current_ir_graph, blk,
1858 new_r_Const_long(current_ir_graph, blk, mode, 1),
1861 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1864 ir_node *blk = get_irn_n(n, -1);
1866 get_irn_dbg_info(n), current_ir_graph, blk,
1869 get_irn_dbg_info(n), current_ir_graph, blk,
1871 new_r_Const_long(current_ir_graph, blk, mode, 1),
1874 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1882 * Do the AddSub optimization, then Transform
1883 * Sub(0,a) -> Minus(a)
1884 * Sub(Mul(a, x), a) -> Mul(a, x-1)
1886 static ir_node *transform_node_Sub(ir_node *n)
1892 n = transform_node_AddSub(n);
1894 mode = get_irn_mode(n);
1895 a = get_Sub_left(n);
1896 b = get_Sub_right(n);
1897 if (mode_is_num(mode) && (classify_Const(a) == CNST_NULL)) {
1899 get_irn_dbg_info(n),
1904 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
1906 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1907 else if (get_opt_reassociation() && get_irn_op(a) == op_Mul) {
1908 ir_node *ma = get_Mul_left(a);
1909 ir_node *mb = get_Mul_right(a);
1912 ir_node *blk = get_irn_n(n, -1);
1914 get_irn_dbg_info(n),
1915 current_ir_graph, blk,
1918 get_irn_dbg_info(n),
1919 current_ir_graph, blk,
1921 new_r_Const_long(current_ir_graph, blk, mode, 1),
1924 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
1927 ir_node *blk = get_irn_n(n, -1);
1929 get_irn_dbg_info(n),
1930 current_ir_graph, blk,
1933 get_irn_dbg_info(n),
1934 current_ir_graph, blk,
1936 new_r_Const_long(current_ir_graph, blk, mode, 1),
1939 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
1947 * Transform Mul(a,-1) into -a.
1948 * Do architecture dependent optimizations on Mul nodes
1950 static ir_node *transform_node_Mul(ir_node *n) {
1952 ir_mode *mode = get_irn_mode(n);
1954 if (mode_is_signed(mode)) {
1956 ir_node *a = get_Mul_left(n);
1957 ir_node *b = get_Mul_right(n);
1959 if (value_of(a) == get_mode_minus_one(mode))
1961 else if (value_of(b) == get_mode_minus_one(mode))
1964 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
1965 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
1969 return arch_dep_replace_mul_with_shifts(n);
1973 * transform a Div Node
1975 static ir_node *transform_node_Div(ir_node *n)
1977 tarval *tv = value_of(n);
1980 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
1982 if (tv != tarval_bad) {
1983 value = new_Const(get_tarval_mode(tv), tv);
1985 DBG_OPT_CSTEVAL(n, value);
1987 else /* Try architecture dependent optimization */
1988 value = arch_dep_replace_div_by_const(n);
1991 /* Turn Div into a tuple (mem, bad, value) */
1992 ir_node *mem = get_Div_mem(n);
1994 turn_into_tuple(n, pn_Div_max);
1995 set_Tuple_pred(n, pn_Div_M, mem);
1996 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
1997 set_Tuple_pred(n, pn_Div_res, value);
2003 * transform a Mod node
2005 static ir_node *transform_node_Mod(ir_node *n)
2007 tarval *tv = value_of(n);
2010 /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */
2012 if (tv != tarval_bad) {
2013 value = new_Const(get_tarval_mode(tv), tv);
2015 DBG_OPT_CSTEVAL(n, value);
2017 else /* Try architecture dependent optimization */
2018 value = arch_dep_replace_mod_by_const(n);
2021 /* Turn Mod into a tuple (mem, bad, value) */
2022 ir_node *mem = get_Mod_mem(n);
2024 turn_into_tuple(n, pn_Mod_max);
2025 set_Tuple_pred(n, pn_Mod_M, mem);
2026 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2027 set_Tuple_pred(n, pn_Mod_res, value);
2033 * transform a DivMod node
2035 static ir_node *transform_node_DivMod(ir_node *n)
2039 ir_node *a = get_DivMod_left(n);
2040 ir_node *b = get_DivMod_right(n);
2041 ir_mode *mode = get_irn_mode(a);
2042 tarval *ta = value_of(a);
2043 tarval *tb = value_of(b);
2045 if (!(mode_is_int(mode) && mode_is_int(get_irn_mode(b))))
2048 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2050 if (tb != tarval_bad) {
2051 if (tb == get_mode_one(get_tarval_mode(tb))) {
2052 b = new_Const (mode, get_mode_null(mode));
2055 DBG_OPT_CSTEVAL(n, b);
2057 else if (ta != tarval_bad) {
2058 tarval *resa, *resb;
2059 resa = tarval_div (ta, tb);
2060 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2061 Jmp for X result!? */
2062 resb = tarval_mod (ta, tb);
2063 if (resb == tarval_bad) return n; /* Causes exception! */
2064 a = new_Const (mode, resa);
2065 b = new_Const (mode, resb);
2068 DBG_OPT_CSTEVAL(n, a);
2069 DBG_OPT_CSTEVAL(n, b);
2071 else { /* Try architecture dependent optimization */
2072 arch_dep_replace_divmod_by_const(&a, &b, n);
2073 evaluated = a != NULL;
2075 } else if (ta == get_mode_null(mode)) {
2076 /* 0 / non-Const = 0 */
2081 if (evaluated) { /* replace by tuple */
2082 ir_node *mem = get_DivMod_mem(n);
2083 turn_into_tuple(n, pn_DivMod_max);
2084 set_Tuple_pred(n, pn_DivMod_M, mem);
2085 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
2086 set_Tuple_pred(n, pn_DivMod_res_div, a);
2087 set_Tuple_pred(n, pn_DivMod_res_mod, b);
2094 * Optimize Abs(x) into x if x is Confirmed >= 0
2095 * Optimize Abs(x) into -x if x is Confirmed <= 0
2097 static ir_node *transform_node_Abs(ir_node *n)
2100 ir_node *a = get_Abs_op(n);
2101 value_classify sign = classify_value_sign(a);
2103 if (sign == VALUE_NEGATIVE) {
2104 ir_mode *mode = get_irn_mode(n);
2107 * We can replace the Abs by -x here.
2108 * We even could add a new Confirm here.
2110 * Note that -x would create a new node, so we could
2111 * not run it in the equivalent_node() context.
2113 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
2114 get_irn_n(n, -1), a, mode);
2116 DBG_OPT_CONFIRM(oldn, n);
2118 else if (sign == VALUE_POSITIVE) {
2119 /* n is positive, Abs is not needed */
2122 DBG_OPT_CONFIRM(oldn, n);
2129 * transform a Cond node
2131 static ir_node *transform_node_Cond(ir_node *n)
2133 /* Replace the Cond by a Jmp if it branches on a constant
2136 ir_node *a = get_Cond_selector(n);
2137 tarval *ta = value_of(a);
2139 /* we need block info which is not available in floating irgs */
2140 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
2143 if ((ta != tarval_bad) &&
2144 (get_irn_mode(a) == mode_b) &&
2145 (get_opt_unreachable_code())) {
2146 /* It's a boolean Cond, branching on a boolean constant.
2147 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2148 jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n));
2149 turn_into_tuple(n, pn_Cond_max);
2150 if (ta == tarval_b_true) {
2151 set_Tuple_pred(n, pn_Cond_false, new_Bad());
2152 set_Tuple_pred(n, pn_Cond_true, jmp);
2154 set_Tuple_pred(n, pn_Cond_false, jmp);
2155 set_Tuple_pred(n, pn_Cond_true, new_Bad());
2157 /* We might generate an endless loop, so keep it alive. */
2158 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
2166 static ir_node *transform_node_Eor(ir_node *n)
2169 ir_node *a = get_Eor_left(n);
2170 ir_node *b = get_Eor_right(n);
2171 ir_mode *mode = get_irn_mode(n);
2175 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
2176 mode, get_mode_null(mode));
2177 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
2179 else if ((mode == mode_b)
2180 && (get_irn_op(a) == op_Proj)
2181 && (get_irn_mode(a) == mode_b)
2182 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)
2183 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2184 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
2185 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2186 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
2188 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
2190 else if ((mode == mode_b)
2191 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) {
2192 /* The Eor is a Not. Replace it by a Not. */
2193 /* ????!!!Extend to bitfield 1111111. */
2194 n = new_r_Not(current_ir_graph, get_irn_n(n, -1), a, mode_b);
2196 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2203 * Transform a boolean Not.
2205 static ir_node *transform_node_Not(ir_node *n)
2208 ir_node *a = get_Not_op(n);
2210 if ( (get_irn_mode(n) == mode_b)
2211 && (get_irn_op(a) == op_Proj)
2212 && (get_irn_mode(a) == mode_b)
2213 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2214 /* We negate a Cmp. The Cmp has the negated result anyways! */
2215 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2216 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
2217 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
2224 * Transform a Cast_type(Const) into a new Const_type
2226 static ir_node *transform_node_Cast(ir_node *n) {
2228 ir_node *pred = get_Cast_op(n);
2229 ir_type *tp = get_irn_type(n);
2231 if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) {
2232 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
2233 get_Const_tarval(pred), tp);
2234 DBG_OPT_CSTEVAL(oldn, n);
2235 } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) {
2236 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred),
2237 get_SymConst_kind(pred), tp);
2238 DBG_OPT_CSTEVAL(oldn, n);
2245 * Transform a Proj(Div) with a non-zero value.
2246 * Removes the exceptions and routes the memory to the NoMem node.
2248 static ir_node *transform_node_Proj_Div(ir_node *proj)
2250 ir_node *n = get_Proj_pred(proj);
2251 ir_node *b = get_Div_right(n);
2254 if (value_not_zero(b)) {
2255 /* div(x, y) && y != 0 */
2256 proj_nr = get_Proj_proj(proj);
2258 /* this node may float if it did not depend on a Confirm */
2259 set_irn_pinned(n, op_pin_state_floats);
2261 if (proj_nr == pn_Div_X_except) {
2262 /* we found an exception handler, remove it */
2263 DBG_OPT_EXC_REM(proj);
2266 else if (proj_nr == pn_Div_M) {
2267 ir_node *res = get_Div_mem(n);
2268 /* the memory Proj can only be removed if we divide by a
2269 real constant, but the node never produce a new memory */
2270 if (value_of(b) != tarval_bad) {
2271 /* this is a Div by a const, we can remove the memory edge */
2272 set_Div_mem(n, get_irg_no_mem(current_ir_graph));
2281 * Transform a Proj(Mod) with a non-zero value.
2282 * Removes the exceptions and routes the memory to the NoMem node.
2284 static ir_node *transform_node_Proj_Mod(ir_node *proj)
2286 ir_node *n = get_Proj_pred(proj);
2287 ir_node *b = get_Mod_right(n);
2290 if (value_not_zero(b)) {
2291 /* mod(x, y) && y != 0 */
2292 proj_nr = get_Proj_proj(proj);
2294 /* this node may float if it did not depend on a Confirm */
2295 set_irn_pinned(n, op_pin_state_floats);
2297 if (proj_nr == pn_Mod_X_except) {
2298 /* we found an exception handler, remove it */
2299 DBG_OPT_EXC_REM(proj);
2301 } else if (proj_nr == pn_Mod_M) {
2302 ir_node *res = get_Mod_mem(n);
2303 /* the memory Proj can only be removed if we divide by a
2304 real constant, but the node never produce a new memory */
2305 if (value_of(b) != tarval_bad) {
2306 /* this is a Mod by a const, we can remove the memory edge */
2307 set_Mod_mem(n, get_irg_no_mem(current_ir_graph));
2311 else if (proj_nr == pn_Mod_res && get_Mod_left(n) == b) {
2312 /* a % a = 0 if a != 0 */
2313 ir_mode *mode = get_irn_mode(proj);
2314 ir_node *res = new_Const(mode, get_mode_null(mode));
2316 DBG_OPT_CSTEVAL(n, res);
2324 * Transform a Proj(DivMod) with a non-zero value.
2325 * Removes the exceptions and routes the memory to the NoMem node.
2327 static ir_node *transform_node_Proj_DivMod(ir_node *proj)
2329 ir_node *n = get_Proj_pred(proj);
2330 ir_node *b = get_DivMod_right(n);
2333 if (value_not_zero(b)) {
2334 /* DivMod(x, y) && y != 0 */
2335 proj_nr = get_Proj_proj(proj);
2337 /* this node may float if it did not depend on a Confirm */
2338 set_irn_pinned(n, op_pin_state_floats);
2340 if (proj_nr == pn_DivMod_X_except) {
2341 /* we found an exception handler, remove it */
2342 DBG_OPT_EXC_REM(proj);
2345 else if (proj_nr == pn_DivMod_M) {
2346 ir_node *res = get_DivMod_mem(n);
2347 /* the memory Proj can only be removed if we divide by a
2348 real constant, but the node never produce a new memory */
2349 if (value_of(b) != tarval_bad) {
2350 /* this is a DivMod by a const, we can remove the memory edge */
2351 set_DivMod_mem(n, get_irg_no_mem(current_ir_graph));
2355 else if (proj_nr == pn_DivMod_res_mod && get_DivMod_left(n) == b) {
2356 /* a % a = 0 if a != 0 */
2357 ir_mode *mode = get_irn_mode(proj);
2358 ir_node *res = new_Const(mode, get_mode_null(mode));
2360 DBG_OPT_CSTEVAL(n, res);
2368 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
2370 static ir_node *transform_node_Proj_Cond(ir_node *proj)
2372 if (get_opt_unreachable_code()) {
2373 ir_node *n = get_Proj_pred(proj);
2374 ir_node *b = get_Cond_selector(n);
2376 if (mode_is_int(get_irn_mode(b))) {
2377 tarval *tb = value_of(b);
2379 if (tb != tarval_bad) {
2380 /* we have a constant switch */
2381 long num = get_Proj_proj(proj);
2383 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
2384 if (get_tarval_long(tb) == num) {
2385 /* Do NOT create a jump here, or we will have 2 control flow ops
2386 * in a block. This case is optimized away in optimize_cf(). */
2390 /* this case will NEVER be taken, kill it */
2401 * Normalizes and optimizes Cmp nodes.
2403 static ir_node *transform_node_Proj_Cmp(ir_node *proj)
2405 if (get_opt_reassociation()) {
2406 ir_node *n = get_Proj_pred(proj);
2407 ir_node *left = get_Cmp_left(n);
2408 ir_node *right = get_Cmp_right(n);
2412 ir_mode *mode = NULL;
2413 long proj_nr = get_Proj_proj(proj);
2416 * First step: normalize the compare op
2417 * by placing the constant on the right site
2418 * or moving the lower address node to the left.
2419 * We ignore the case that both are constants
2420 * this case should be optimized away.
2422 if (get_irn_op(right) == op_Const)
2424 else if (get_irn_op(left) == op_Const) {
2429 proj_nr = get_inversed_pnc(proj_nr);
2432 else if (get_irn_idx(left) > get_irn_idx(right)) {
2438 proj_nr = get_inversed_pnc(proj_nr);
2443 * Second step: Try to reduce the magnitude
2444 * of a constant. This may help to generate better code
2445 * later and may help to normalize more compares.
2446 * Of course this is only possible for integer values.
2449 mode = get_irn_mode(c);
2450 tv = get_Const_tarval(c);
2452 if (tv != tarval_bad) {
2453 /* the following optimization is possible on modes without Overflow
2454 * on Unary Minus or on == and !=:
2455 * -a CMP c ==> a swap(CMP) -c
2457 * Beware: for two-complement Overflow may occur, so only == and != can
2458 * be optimized, see this:
2459 * -MININT < 0 =/=> MININT > 0 !!!
2461 if (get_opt_constant_folding() && get_irn_op(left) == op_Minus &&
2462 (!mode_overflow_on_unary_Minus(mode) ||
2463 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
2464 left = get_Minus_op(left);
2465 tv = tarval_sub(get_mode_null(mode), tv);
2467 proj_nr = get_inversed_pnc(proj_nr);
2471 /* for integer modes, we have more */
2472 if (mode_is_int(mode)) {
2473 /* Ne includes Unordered which is not possible on integers.
2474 * However, frontends often use this wrong, so fix it here */
2475 if (proj_nr & pn_Cmp_Uo) {
2476 proj_nr &= ~pn_Cmp_Uo;
2477 set_Proj_proj(proj, proj_nr);
2480 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
2481 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
2482 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
2483 tv = tarval_sub(tv, get_mode_one(mode));
2485 proj_nr ^= pn_Cmp_Eq;
2488 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
2489 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
2490 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
2491 tv = tarval_add(tv, get_mode_one(mode));
2493 proj_nr ^= pn_Cmp_Eq;
2497 /* the following reassociations work only for == and != */
2498 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
2500 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
2501 if (classify_tarval(tv) == TV_CLASSIFY_NULL && get_irn_op(left) == op_Sub) {
2502 right = get_Sub_right(left);
2503 left = get_Sub_left(left);
2505 tv = value_of(right);
2509 if (tv != tarval_bad) {
2510 ir_op *op = get_irn_op(left);
2512 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
2514 ir_node *c1 = get_Sub_right(left);
2515 tarval *tv2 = value_of(c1);
2517 if (tv2 != tarval_bad) {
2518 tv2 = tarval_add(tv, value_of(c1));
2520 if (tv2 != tarval_bad) {
2521 left = get_Sub_left(left);
2527 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
2528 else if (op == op_Add) {
2529 ir_node *a_l = get_Add_left(left);
2530 ir_node *a_r = get_Add_right(left);
2534 if (get_irn_op(a_l) == op_Const) {
2536 tv2 = value_of(a_l);
2540 tv2 = value_of(a_r);
2543 if (tv2 != tarval_bad) {
2544 tv2 = tarval_sub(tv, tv2);
2546 if (tv2 != tarval_bad) {
2553 /* -a == c ==> a == -c, -a != c ==> a != -c */
2554 else if (op == op_Minus) {
2555 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
2557 if (tv2 != tarval_bad) {
2558 left = get_Minus_op(left);
2565 /* the following reassociations work only for <= */
2566 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2567 if (tv != tarval_bad) {
2568 ir_op *op = get_irn_op(left);
2570 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
2578 * optimization for AND:
2580 * And(x, C) == C ==> And(x, C) != 0
2581 * And(x, C) != C ==> And(X, C) == 0
2583 * if C is a single Bit constant.
2585 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) &&
2586 (get_irn_op(left) == op_And)) {
2587 if (is_single_bit_tarval(tv)) {
2588 /* check for Constant's match. We have check hare the tarvals,
2589 because our const might be changed */
2590 ir_node *la = get_And_left(left);
2591 ir_node *ra = get_And_right(left);
2592 if ((is_Const(la) && get_Const_tarval(la) == tv) ||
2593 (is_Const(ra) && get_Const_tarval(ra) == tv)) {
2594 /* fine: do the transformation */
2595 tv = get_mode_null(get_tarval_mode(tv));
2596 proj_nr ^= pn_Cmp_Leg;
2601 } /* tarval != bad */
2605 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
2607 if (changed & 2) /* need a new Const */
2608 right = new_Const(mode, tv);
2610 /* create a new compare */
2611 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block,
2614 set_Proj_pred(proj, n);
2615 set_Proj_proj(proj, proj_nr);
2622 * Does all optimizations on nodes that must be done on it's Proj's
2623 * because of creating new nodes.
2625 static ir_node *transform_node_Proj(ir_node *proj)
2627 ir_node *n = get_Proj_pred(proj);
2629 switch (get_irn_opcode(n)) {
2631 return transform_node_Proj_Div(proj);
2634 return transform_node_Proj_Mod(proj);
2637 return transform_node_Proj_DivMod(proj);
2640 return transform_node_Proj_Cond(proj);
2643 return transform_node_Proj_Cmp(proj);
2646 /* should not happen, but if it does will be optimized away */
2647 return equivalent_node_Proj(proj);
2656 * Move Confirms down through Phi nodes.
2658 static ir_node *transform_node_Phi(ir_node *phi) {
2660 ir_mode *mode = get_irn_mode(phi);
2662 if (mode_is_reference(mode)) {
2663 n = get_irn_arity(phi);
2665 /* Beware of Phi0 */
2667 ir_node *pred = get_irn_n(phi, 0);
2668 ir_node *bound, *new_Phi, *block, **in;
2671 if (! is_Confirm(pred))
2674 bound = get_Confirm_bound(pred);
2675 pnc = get_Confirm_cmp(pred);
2677 NEW_ARR_A(ir_node *, in, n);
2678 in[0] = get_Confirm_value(pred);
2680 for (i = 1; i < n; ++i) {
2681 pred = get_irn_n(phi, i);
2683 if (! is_Confirm(pred) ||
2684 get_Confirm_bound(pred) != bound ||
2685 get_Confirm_cmp(pred) != pnc)
2687 in[i] = get_Confirm_value(pred);
2689 /* move the Confirm nodes "behind" the Phi */
2690 block = get_irn_n(phi, -1);
2691 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
2692 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
2699 * returns the operands of a commutative bin-op, if one operand is
2700 * a const, it is returned as the second one.
2702 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c)
2704 ir_node *op_a = get_binop_left(binop);
2705 ir_node *op_b = get_binop_right(binop);
2707 assert(is_op_commutative(get_irn_op(binop)));
2709 if (get_irn_op(op_a) == op_Const) {
2720 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
2721 * Such pattern may arise in bitfield stores.
2723 * value c4 value c4 & c2
2724 * AND c3 AND c1 | c3
2729 static ir_node *transform_node_Or_bf_store(ir_node *or)
2733 ir_node *and_l, *c3;
2734 ir_node *value, *c4;
2735 ir_node *new_and, *new_const, *block;
2736 ir_mode *mode = get_irn_mode(or);
2738 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
2740 get_comm_Binop_Ops(or, &and, &c1);
2741 if ((get_irn_op(c1) != op_Const) || (get_irn_op(and) != op_And))
2744 get_comm_Binop_Ops(and, &or_l, &c2);
2745 if ((get_irn_op(c2) != op_Const) || (get_irn_op(or_l) != op_Or))
2748 get_comm_Binop_Ops(or_l, &and_l, &c3);
2749 if ((get_irn_op(c3) != op_Const) || (get_irn_op(and_l) != op_And))
2752 get_comm_Binop_Ops(and_l, &value, &c4);
2753 if (get_irn_op(c4) != op_Const)
2756 /* ok, found the pattern, check for conditions */
2757 assert(mode == get_irn_mode(and));
2758 assert(mode == get_irn_mode(or_l));
2759 assert(mode == get_irn_mode(and_l));
2761 tv1 = get_Const_tarval(c1);
2762 tv2 = get_Const_tarval(c2);
2763 tv3 = get_Const_tarval(c3);
2764 tv4 = get_Const_tarval(c4);
2766 tv = tarval_or(tv4, tv2);
2767 if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) {
2768 /* have at least one 0 at the same bit position */
2772 n_tv4 = tarval_not(tv4);
2773 if (tv3 != tarval_and(tv3, n_tv4)) {
2774 /* bit in the or_mask is outside the and_mask */
2778 n_tv2 = tarval_not(tv2);
2779 if (tv1 != tarval_and(tv1, n_tv2)) {
2780 /* bit in the or_mask is outside the and_mask */
2784 /* ok, all conditions met */
2785 block = get_irn_n(or, -1);
2787 new_and = new_r_And(current_ir_graph, block,
2788 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
2790 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
2792 set_Or_left(or, new_and);
2793 set_Or_right(or, new_const);
2795 /* check for more */
2796 return transform_node_Or_bf_store(or);
2800 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
2802 static ir_node *transform_node_Or_Rot(ir_node *or)
2804 ir_mode *mode = get_irn_mode(or);
2805 ir_node *shl, *shr, *block;
2806 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
2809 if (! mode_is_int(mode))
2812 shl = get_binop_left(or);
2813 shr = get_binop_right(or);
2815 if (get_irn_op(shl) == op_Shr) {
2816 if (get_irn_op(shr) != op_Shl)
2823 else if (get_irn_op(shl) != op_Shl)
2825 else if (get_irn_op(shr) != op_Shr)
2828 x = get_Shl_left(shl);
2829 if (x != get_Shr_left(shr))
2832 c1 = get_Shl_right(shl);
2833 c2 = get_Shr_right(shr);
2834 if (get_irn_op(c1) == op_Const && get_irn_op(c2) == op_Const) {
2835 tv1 = get_Const_tarval(c1);
2836 if (! tarval_is_long(tv1))
2839 tv2 = get_Const_tarval(c2);
2840 if (! tarval_is_long(tv2))
2843 if (get_tarval_long(tv1) + get_tarval_long(tv2)
2844 != get_mode_size_bits(mode))
2847 /* yet, condition met */
2848 block = get_irn_n(or, -1);
2850 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
2852 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
2855 else if (get_irn_op(c1) == op_Sub) {
2859 if (get_Sub_right(sub) != v)
2862 c1 = get_Sub_left(sub);
2863 if (get_irn_op(c1) != op_Const)
2866 tv1 = get_Const_tarval(c1);
2867 if (! tarval_is_long(tv1))
2870 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
2873 /* yet, condition met */
2874 block = get_nodes_block(or);
2876 /* a Rot right is not supported, so use a rot left */
2877 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
2879 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
2882 else if (get_irn_op(c2) == op_Sub) {
2886 c1 = get_Sub_left(sub);
2887 if (get_irn_op(c1) != op_Const)
2890 tv1 = get_Const_tarval(c1);
2891 if (! tarval_is_long(tv1))
2894 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
2897 /* yet, condition met */
2898 block = get_irn_n(or, -1);
2901 n = new_r_Rot(current_ir_graph, block, x, v, mode);
2903 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
2913 static ir_node *transform_node_Or(ir_node *or)
2915 or = transform_node_Or_bf_store(or);
2916 or = transform_node_Or_Rot(or);
2922 static ir_node *transform_node(ir_node *n);
2925 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl.
2927 * Should be moved to reassociation?
2929 static ir_node *transform_node_shift(ir_node *n)
2931 ir_node *left, *right;
2932 tarval *tv1, *tv2, *res;
2934 int modulo_shf, flag;
2936 left = get_binop_left(n);
2938 /* different operations */
2939 if (get_irn_op(left) != get_irn_op(n))
2942 right = get_binop_right(n);
2943 tv1 = value_of(right);
2944 if (tv1 == tarval_bad)
2947 tv2 = value_of(get_binop_right(left));
2948 if (tv2 == tarval_bad)
2951 res = tarval_add(tv1, tv2);
2953 /* beware: a simple replacement works only, if res < modulo shift */
2954 mode = get_irn_mode(n);
2958 modulo_shf = get_mode_modulo_shift(mode);
2959 if (modulo_shf > 0) {
2960 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
2962 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
2969 /* ok, we can replace it */
2970 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
2972 in[0] = get_binop_left(left);
2973 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
2975 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
2977 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
2979 return transform_node(irn);
2984 #define transform_node_Shr transform_node_shift
2985 #define transform_node_Shrs transform_node_shift
2986 #define transform_node_Shl transform_node_shift
2989 * Remove dead blocks and nodes in dead blocks
2990 * in keep alive list. We do not generate a new End node.
2992 static ir_node *transform_node_End(ir_node *n) {
2993 int i, n_keepalives = get_End_n_keepalives(n);
2995 for (i = 0; i < n_keepalives; ++i) {
2996 ir_node *ka = get_End_keepalive(n, i);
2998 if (is_Block_dead(ka)) {
2999 set_End_keepalive(n, i, new_Bad());
3002 else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka)))
3003 set_End_keepalive(n, i, new_Bad());
3009 * Optimize a Mux into some simpler cases.
3011 static ir_node *transform_node_Mux(ir_node *n)
3013 ir_node *oldn = n, *sel = get_Mux_sel(n);
3014 ir_mode *mode = get_irn_mode(n);
3016 if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(mode)) {
3017 ir_node *cmp = get_Proj_pred(sel);
3018 long proj_nr = get_Proj_proj(sel);
3019 ir_node *f = get_Mux_false(n);
3020 ir_node *t = get_Mux_true(n);
3022 if (get_irn_op(cmp) == op_Cmp && classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
3023 ir_node *block = get_irn_n(n, -1);
3026 * Note: normalization puts the constant on the right site,
3027 * so we check only one case.
3029 * Note further that these optimization work even for floating point
3030 * with NaN's because -NaN == NaN.
3031 * However, if +0 and -0 is handled differently, we cannot use the first one.
3033 if (get_irn_op(f) == op_Minus &&
3034 get_Minus_op(f) == t &&
3035 get_Cmp_left(cmp) == t) {
3037 if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3038 /* Mux(a >=/> 0, -a, a) ==> Abs(a) */
3039 n = new_rd_Abs(get_irn_dbg_info(n),
3043 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3046 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3047 /* Mux(a <=/< 0, -a, a) ==> Minus(Abs(a)) */
3048 n = new_rd_Abs(get_irn_dbg_info(n),
3052 n = new_rd_Minus(get_irn_dbg_info(n),
3057 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3061 else if (get_irn_op(t) == op_Minus &&
3062 get_Minus_op(t) == f &&
3063 get_Cmp_left(cmp) == f) {
3065 if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3066 /* Mux(a <=/< 0, a, -a) ==> Abs(a) */
3067 n = new_rd_Abs(get_irn_dbg_info(n),
3071 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3074 else if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3075 /* Mux(a >=/> 0, a, -a) ==> Minus(Abs(a)) */
3076 n = new_rd_Abs(get_irn_dbg_info(n),
3080 n = new_rd_Minus(get_irn_dbg_info(n),
3085 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3090 if (mode_is_int(mode) && mode_is_signed(mode) &&
3091 get_mode_arithmetic(mode) == irma_twos_complement) {
3092 ir_node *x = get_Cmp_left(cmp);
3094 /* the following optimization works only with signed integer two-complement mode */
3096 if (mode == get_irn_mode(x)) {
3098 * FIXME: this restriction is two rigid, as it would still
3099 * work if mode(x) = Hs and mode == Is, but at least it removes
3102 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Le) &&
3103 classify_Const(t) == CNST_ALL_ONE &&
3104 classify_Const(f) == CNST_NULL) {
3106 * Mux(x:T </<= 0, 0, -1) -> Shrs(x, sizeof_bits(T) - 1)
3110 n = new_rd_Shrs(get_irn_dbg_info(n),
3111 current_ir_graph, block, x,
3112 new_r_Const_long(current_ir_graph, block, mode_Iu,
3113 get_mode_size_bits(mode) - 1),
3115 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3118 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) &&
3119 classify_Const(t) == CNST_ONE &&
3120 classify_Const(f) == CNST_NULL) {
3122 * Mux(x:T >/>= 0, 0, 1) -> Shr(-x, sizeof_bits(T) - 1)
3126 n = new_rd_Shr(get_irn_dbg_info(n),
3127 current_ir_graph, block,
3128 new_r_Minus(current_ir_graph, block, x, mode),
3129 new_r_Const_long(current_ir_graph, block, mode_Iu,
3130 get_mode_size_bits(mode) - 1),
3132 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3139 return arch_transform_node_Mux(n);
3143 * Optimize a Psi into some simpler cases.
3145 static ir_node *transform_node_Psi(ir_node *n) {
3147 return transform_node_Mux(n);
3153 * Tries several [inplace] [optimizing] transformations and returns an
3154 * equivalent node. The difference to equivalent_node() is that these
3155 * transformations _do_ generate new nodes, and thus the old node must
3156 * not be freed even if the equivalent node isn't the old one.
3158 static ir_node *transform_node(ir_node *n)
3160 if (n->op->ops.transform_node)
3161 n = n->op->ops.transform_node(n);
3166 * sSets the default transform node operation for an ir_op_ops.
3168 * @param code the opcode for the default operation
3169 * @param ops the operations initialized
3174 static ir_op_ops *firm_set_default_transform_node(opcode code, ir_op_ops *ops)
3178 ops->transform_node = transform_node_##a; \
3212 /* **************** Common Subexpression Elimination **************** */
3214 /** The size of the hash table used, should estimate the number of nodes
3216 #define N_IR_NODES 512
3218 /** Compares the attributes of two Const nodes. */
3219 static int node_cmp_attr_Const(ir_node *a, ir_node *b)
3221 return (get_Const_tarval(a) != get_Const_tarval(b))
3222 || (get_Const_type(a) != get_Const_type(b));
3225 /** Compares the attributes of two Proj nodes. */
3226 static int node_cmp_attr_Proj(ir_node *a, ir_node *b)
3228 return get_irn_proj_attr (a) != get_irn_proj_attr (b);
3231 /** Compares the attributes of two Filter nodes. */
3232 static int node_cmp_attr_Filter(ir_node *a, ir_node *b)
3234 return get_Filter_proj(a) != get_Filter_proj(b);
3237 /** Compares the attributes of two Alloc nodes. */
3238 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b)
3240 return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where)
3241 || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type);
3244 /** Compares the attributes of two Free nodes. */
3245 static int node_cmp_attr_Free(ir_node *a, ir_node *b)
3247 return (get_irn_free_attr(a).where != get_irn_free_attr(b).where)
3248 || (get_irn_free_attr(a).type != get_irn_free_attr(b).type);
3251 /** Compares the attributes of two SymConst nodes. */
3252 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b)
3254 return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num)
3255 || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p)
3256 || (get_irn_symconst_attr(a).tp != get_irn_symconst_attr(b).tp);
3259 /** Compares the attributes of two Call nodes. */
3260 static int node_cmp_attr_Call(ir_node *a, ir_node *b)
3262 return (get_irn_call_attr(a) != get_irn_call_attr(b));
3265 /** Compares the attributes of two Sel nodes. */
3266 static int node_cmp_attr_Sel(ir_node *a, ir_node *b)
3268 return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind)
3269 || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name)
3270 || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner)
3271 || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name)
3272 || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type);
3275 /** Compares the attributes of two Phi nodes. */
3276 static int node_cmp_attr_Phi(ir_node *a, ir_node *b)
3278 return get_irn_phi_attr (a) != get_irn_phi_attr (b);
3281 /** Compares the attributes of two Cast nodes. */
3282 static int node_cmp_attr_Cast(ir_node *a, ir_node *b)
3284 return get_Cast_type(a) != get_Cast_type(b);
3287 /** Compares the attributes of two Load nodes. */
3288 static int node_cmp_attr_Load(ir_node *a, ir_node *b)
3290 if (get_Load_volatility(a) == volatility_is_volatile ||
3291 get_Load_volatility(b) == volatility_is_volatile)
3292 /* NEVER do CSE on volatile Loads */
3295 return get_Load_mode(a) != get_Load_mode(b);
3298 /** Compares the attributes of two Store nodes. */
3299 static int node_cmp_attr_Store(ir_node *a, ir_node *b)
3301 /* NEVER do CSE on volatile Stores */
3302 return (get_Store_volatility(a) == volatility_is_volatile ||
3303 get_Store_volatility(b) == volatility_is_volatile);
3306 /** Compares the attributes of two Confirm nodes. */
3307 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b)
3309 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
3313 * Set the default node attribute compare operation for an ir_op_ops.
3315 * @param code the opcode for the default operation
3316 * @param ops the operations initialized
3321 static ir_op_ops *firm_set_default_node_cmp_attr(opcode code, ir_op_ops *ops)
3325 ops->node_cmp_attr = node_cmp_attr_##a; \
3351 * Compare function for two nodes in the hash table. Gets two
3352 * nodes as parameters. Returns 0 if the nodes are a cse.
3354 int identities_cmp(const void *elt, const void *key)
3362 if (a == b) return 0;
3364 if ((get_irn_op(a) != get_irn_op(b)) ||
3365 (get_irn_mode(a) != get_irn_mode(b))) return 1;
3367 /* compare if a's in and b's in are of equal length */
3368 irn_arity_a = get_irn_intra_arity (a);
3369 if (irn_arity_a != get_irn_intra_arity(b))
3372 /* for block-local cse and op_pin_state_pinned nodes: */
3373 if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) {
3374 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
3378 /* compare a->in[0..ins] with b->in[0..ins] */
3379 for (i = 0; i < irn_arity_a; i++)
3380 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
3384 * here, we already now that the nodes are identical except their
3387 if (a->op->ops.node_cmp_attr)
3388 return a->op->ops.node_cmp_attr(a, b);
3394 * Calculate a hash value of a node.
3397 ir_node_hash (ir_node *node)
3402 if (node->op == op_Const) {
3403 /* special value for const, as they only differ in their tarval. */
3404 h = HASH_PTR(node->attr.con.tv);
3405 h = 9*h + HASH_PTR(get_irn_mode(node));
3406 } else if (node->op == op_SymConst) {
3407 /* special value for const, as they only differ in their symbol. */
3408 h = HASH_PTR(node->attr.i.sym.type_p);
3409 h = 9*h + HASH_PTR(get_irn_mode(node));
3412 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
3413 h = irn_arity = get_irn_intra_arity(node);
3415 /* consider all in nodes... except the block if not a control flow. */
3416 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
3417 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
3421 h = 9*h + HASH_PTR(get_irn_mode(node));
3423 h = 9*h + HASH_PTR(get_irn_op(node));
3430 new_identities(void) {
3431 return new_pset(identities_cmp, N_IR_NODES);
3435 del_identities(pset *value_table) {
3436 del_pset(value_table);
3440 * Return the canonical node computing the same value as n.
3441 * Looks up the node in a hash table.
3443 * For Const nodes this is performed in the constructor, too. Const
3444 * nodes are extremely time critical because of their frequent use in
3445 * constant string arrays.
3447 static INLINE ir_node *identify(pset *value_table, ir_node *n)
3451 if (!value_table) return n;
3453 if (get_opt_reassociation()) {
3454 if (is_op_commutative(get_irn_op(n))) {
3455 ir_node *l = get_binop_left(n);
3456 ir_node *r = get_binop_right(n);
3458 /* for commutative operators perform a OP b == b OP a */
3459 if (get_irn_idx(l) > get_irn_idx(r)) {
3460 set_binop_left(n, r);
3461 set_binop_right(n, l);
3466 o = pset_find(value_table, n, ir_node_hash (n));
3475 * During construction we set the op_pin_state_pinned flag in the graph right when the
3476 * optimization is performed. The flag turning on procedure global cse could
3477 * be changed between two allocations. This way we are safe.
3479 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
3482 n = identify(value_table, n);
3483 if (get_irn_n(old, -1) != get_irn_n(n, -1))
3484 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3489 * Return the canonical node computing the same value as n.
3490 * Looks up the node in a hash table, enters it in the table
3491 * if it isn't there yet.
3493 ir_node *identify_remember(pset *value_table, ir_node *n)
3497 if (!value_table) return n;
3499 if (get_opt_reassociation()) {
3500 if (is_op_commutative(get_irn_op(n))) {
3501 ir_node *l = get_binop_left(n);
3502 ir_node *r = get_binop_right(n);
3504 /* for commutative operators perform a OP b == b OP a */
3506 set_binop_left(n, r);
3507 set_binop_right(n, l);
3512 /* lookup or insert in hash table with given hash key. */
3513 o = pset_insert (value_table, n, ir_node_hash (n));
3522 /* Add a node to the identities value table. */
3523 void add_identities(pset *value_table, ir_node *node) {
3524 if (get_opt_cse() && is_no_Block(node))
3525 identify_remember(value_table, node);
3528 /* Visit each node in the value table of a graph. */
3529 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
3531 ir_graph *rem = current_ir_graph;
3533 current_ir_graph = irg;
3534 foreach_pset(irg->value_table, node)
3536 current_ir_graph = rem;
3540 * garbage in, garbage out. If a node has a dead input, i.e., the
3541 * Bad node is input to the node, return the Bad node.
3543 static INLINE ir_node *gigo(ir_node *node)
3546 ir_op *op = get_irn_op(node);
3548 /* remove garbage blocks by looking at control flow that leaves the block
3549 and replacing the control flow by Bad. */
3550 if (get_irn_mode(node) == mode_X) {
3551 ir_node *block = get_nodes_block(skip_Proj(node));
3553 /* Don't optimize nodes in immature blocks. */
3554 if (!get_Block_matured(block)) return node;
3555 /* Don't optimize End, may have Bads. */
3556 if (op == op_End) return node;
3558 if (is_Block(block)) {
3559 irn_arity = get_irn_arity(block);
3560 for (i = 0; i < irn_arity; i++) {
3561 if (!is_Bad(get_irn_n(block, i)))
3564 if (i == irn_arity) return new_Bad();
3568 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
3569 blocks predecessors is dead. */
3570 if ( op != op_Block && op != op_Phi && op != op_Tuple) {
3571 irn_arity = get_irn_arity(node);
3574 * Beware: we can only read the block of a non-floating node.
3576 if (is_irn_pinned_in_irg(node) &&
3577 is_Block_dead(get_nodes_block(node)))
3580 for (i = 0; i < irn_arity; i++) {
3581 ir_node *pred = get_irn_n(node, i);
3586 /* Propagating Unknowns here seems to be a bad idea, because
3587 sometimes we need a node as a input and did not want that
3589 However, i might be useful to move this into a later phase
3590 (it you thing optimizing such code is useful). */
3591 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
3592 return new_Unknown(get_irn_mode(node));
3597 /* With this code we violate the agreement that local_optimize
3598 only leaves Bads in Block, Phi and Tuple nodes. */
3599 /* If Block has only Bads as predecessors it's garbage. */
3600 /* If Phi has only Bads as predecessors it's garbage. */
3601 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
3602 irn_arity = get_irn_arity(node);
3603 for (i = 0; i < irn_arity; i++) {
3604 if (!is_Bad(get_irn_n(node, i))) break;
3606 if (i == irn_arity) node = new_Bad();
3613 * These optimizations deallocate nodes from the obstack.
3614 * It can only be called if it is guaranteed that no other nodes
3615 * reference this one, i.e., right after construction of a node.
3617 * current_ir_graph must be set to the graph of the node!
3619 ir_node *optimize_node(ir_node *n)
3623 opcode iro = get_irn_opcode(n);
3625 /* Always optimize Phi nodes: part of the construction. */
3626 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
3628 /* constant expression evaluation / constant folding */
3629 if (get_opt_constant_folding()) {
3630 /* neither constants nor Tuple values can be evaluated */
3631 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
3632 /* try to evaluate */
3633 tv = computed_value(n);
3634 if (tv != tarval_bad) {
3636 ir_type *old_tp = get_irn_type(n);
3637 int i, arity = get_irn_arity(n);
3641 * Try to recover the type of the new expression.
3643 for (i = 0; i < arity && !old_tp; ++i)
3644 old_tp = get_irn_type(get_irn_n(n, i));
3647 * we MUST copy the node here temporary, because it's still needed
3648 * for DBG_OPT_CSTEVAL
3650 node_size = offsetof(ir_node, attr) + n->op->attr_size;
3651 oldn = alloca(node_size);
3653 memcpy(oldn, n, node_size);
3654 CLONE_ARR_A(ir_node *, oldn->in, n->in);
3656 /* ARG, copy the in array, we need it for statistics */
3657 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
3659 /* note the inplace edges module */
3660 edges_node_deleted(n, current_ir_graph);
3662 /* evaluation was successful -- replace the node. */
3663 irg_kill_node(current_ir_graph, n);
3664 nw = new_Const(get_tarval_mode (tv), tv);
3666 if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv))
3667 set_Const_type(nw, old_tp);
3668 DBG_OPT_CSTEVAL(oldn, nw);
3674 /* remove unnecessary nodes */
3675 if (get_opt_constant_folding() ||
3676 (iro == iro_Phi) || /* always optimize these nodes. */
3678 (iro == iro_Proj) ||
3679 (iro == iro_Block) ) /* Flags tested local. */
3680 n = equivalent_node (n);
3682 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
3684 /* Common Subexpression Elimination.
3686 * Checks whether n is already available.
3687 * The block input is used to distinguish different subexpressions. Right
3688 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
3689 * subexpressions within a block.
3692 n = identify_cons (current_ir_graph->value_table, n);
3695 edges_node_deleted(oldn, current_ir_graph);
3697 /* We found an existing, better node, so we can deallocate the old node. */
3698 irg_kill_node(current_ir_graph, oldn);
3702 /* Some more constant expression evaluation that does not allow to
3704 iro = get_irn_opcode(n);
3705 if (get_opt_constant_folding() ||
3706 (iro == iro_Cond) ||
3707 (iro == iro_Proj) ||
3708 (iro == iro_Sel)) /* Flags tested local. */
3709 n = transform_node (n);
3711 /* Remove nodes with dead (Bad) input.
3712 Run always for transformation induced Bads. */
3715 /* Now we have a legal, useful node. Enter it in hash table for CSE */
3716 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
3717 n = identify_remember (current_ir_graph->value_table, n);
3725 * These optimizations never deallocate nodes (in place). This can cause dead
3726 * nodes lying on the obstack. Remove these by a dead node elimination,
3727 * i.e., a copying garbage collection.
3729 ir_node *optimize_in_place_2(ir_node *n)
3733 opcode iro = get_irn_opcode(n);
3735 if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n;
3737 /* constant expression evaluation / constant folding */
3738 if (get_opt_constant_folding()) {
3739 /* neither constants nor Tuple values can be evaluated */
3740 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
3741 /* try to evaluate */
3742 tv = computed_value(n);
3743 if (tv != tarval_bad) {
3744 /* evaluation was successful -- replace the node. */
3745 ir_type *old_tp = get_irn_type(n);
3746 int i, arity = get_irn_arity(n);
3749 * Try to recover the type of the new expression.
3751 for (i = 0; i < arity && !old_tp; ++i)
3752 old_tp = get_irn_type(get_irn_n(n, i));
3754 n = new_Const(get_tarval_mode(tv), tv);
3756 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
3757 set_Const_type(n, old_tp);
3759 DBG_OPT_CSTEVAL(oldn, n);
3765 /* remove unnecessary nodes */
3766 if (get_opt_constant_folding() ||
3767 (iro == iro_Phi) || /* always optimize these nodes. */
3768 (iro == iro_Id) || /* ... */
3769 (iro == iro_Proj) || /* ... */
3770 (iro == iro_Block) ) /* Flags tested local. */
3771 n = equivalent_node(n);
3773 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
3775 /** common subexpression elimination **/
3776 /* Checks whether n is already available. */
3777 /* The block input is used to distinguish different subexpressions. Right
3778 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
3779 subexpressions within a block. */
3780 if (get_opt_cse()) {
3781 n = identify(current_ir_graph->value_table, n);
3784 /* Some more constant expression evaluation. */
3785 iro = get_irn_opcode(n);
3786 if (get_opt_constant_folding() ||
3787 (iro == iro_Cond) ||
3788 (iro == iro_Proj) ||
3789 (iro == iro_Sel)) /* Flags tested local. */
3790 n = transform_node(n);
3792 /* Remove nodes with dead (Bad) input.
3793 Run always for transformation induced Bads. */
3796 /* Now we can verify the node, as it has no dead inputs any more. */
3799 /* Now we have a legal, useful node. Enter it in hash table for cse.
3800 Blocks should be unique anyways. (Except the successor of start:
3801 is cse with the start block!) */
3802 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
3803 n = identify_remember(current_ir_graph->value_table, n);
3809 * Wrapper for external use, set proper status bits after optimization.
3811 ir_node *optimize_in_place(ir_node *n)
3813 /* Handle graph state */
3814 assert(get_irg_phase_state(current_ir_graph) != phase_building);
3816 if (get_opt_global_cse())
3817 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3818 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
3819 set_irg_outs_inconsistent(current_ir_graph);
3821 /* FIXME: Maybe we could also test whether optimizing the node can
3822 change the control graph. */
3823 set_irg_doms_inconsistent(current_ir_graph);
3824 return optimize_in_place_2 (n);
3828 * Sets the default operation for an ir_ops.
3830 ir_op_ops *firm_set_default_operations(opcode code, ir_op_ops *ops)
3832 ops = firm_set_default_computed_value(code, ops);
3833 ops = firm_set_default_equivalent_node(code, ops);
3834 ops = firm_set_default_transform_node(code, ops);
3835 ops = firm_set_default_node_cmp_attr(code, ops);
3836 ops = firm_set_default_get_type(code, ops);
3837 ops = firm_set_default_get_type_attr(code, ops);
3838 ops = firm_set_default_get_entity_attr(code, ops);