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
1278 static ir_node *equivalent_node_Sync(ir_node *n)
1283 ir_node *first_val = NULL; /* to shutup gcc */
1285 if (!get_opt_normalize()) return n;
1287 n_preds = get_Sync_n_preds(n);
1289 /* Find first non-self-referencing input */
1290 for (i = 0; i < n_preds; ++i) {
1291 first_val = get_Sync_pred(n, i);
1292 if ((first_val != n) /* not self pointer */ &&
1293 (! is_Bad(first_val))
1294 ) { /* value not dead */
1295 break; /* then found first value. */
1300 /* A totally Bad or self-referencing Sync (we didn't break the above loop) */
1303 /* search the rest of inputs, determine if any of these
1304 are non-self-referencing */
1305 while (++i < n_preds) {
1306 ir_node *scnd_val = get_Sync_pred(n, i);
1307 if ((scnd_val != n) &&
1308 (scnd_val != first_val) &&
1309 (! is_Bad(scnd_val))
1315 /* Fold, if no multiple distinct non-self-referencing inputs */
1317 DBG_OPT_SYNC(oldn, n);
1323 * optimize Proj(Tuple) and gigo() for ProjX in Bad block,
1324 * ProjX(Load) and ProjX(Store)
1326 static ir_node *equivalent_node_Proj(ir_node *n)
1330 ir_node *a = get_Proj_pred(n);
1332 if ( get_irn_op(a) == op_Tuple) {
1333 /* Remove the Tuple/Proj combination. */
1334 if ( get_Proj_proj(n) <= get_Tuple_n_preds(a) ) {
1335 n = get_Tuple_pred(a, get_Proj_proj(n));
1336 DBG_OPT_TUPLE(oldn, a, n);
1338 assert(0); /* This should not happen! */
1342 else if (get_irn_mode(n) == mode_X) {
1343 if (is_Block_dead(get_nodes_block(skip_Proj(n)))) {
1344 /* Remove dead control flow -- early gigo(). */
1347 else if (get_opt_ldst_only_null_ptr_exceptions()) {
1348 ir_op *op = get_irn_op(a);
1350 if (op == op_Load || op == op_Store) {
1351 /* get the load/store address */
1352 ir_node *addr = get_irn_n(a, 1);
1353 if (value_not_null(addr)) {
1354 /* this node may float if it did not depend on a Confirm */
1355 set_irn_pinned(a, op_pin_state_floats);
1369 static ir_node *equivalent_node_Id(ir_node *n)
1375 } while (get_irn_op(n) == op_Id);
1377 DBG_OPT_ID(oldn, n);
1384 static ir_node *equivalent_node_Mux(ir_node *n)
1386 ir_node *oldn = n, *sel = get_Mux_sel(n);
1387 tarval *ts = value_of(sel);
1389 /* Mux(true, f, t) == t */
1390 if (ts == tarval_b_true) {
1391 n = get_Mux_true(n);
1392 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1394 /* Mux(false, f, t) == f */
1395 else if (ts == tarval_b_false) {
1396 n = get_Mux_false(n);
1397 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1399 /* Mux(v, x, x) == x */
1400 else if (get_Mux_false(n) == get_Mux_true(n)) {
1401 n = get_Mux_true(n);
1402 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1404 else if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(get_irn_mode(n))) {
1405 ir_node *cmp = get_Proj_pred(sel);
1406 long proj_nr = get_Proj_proj(sel);
1407 ir_node *b = get_Mux_false(n);
1408 ir_node *a = get_Mux_true(n);
1411 * Note: normalization puts the constant on the right site,
1412 * so we check only one case.
1414 * Note further that these optimization work even for floating point
1415 * with NaN's because -NaN == NaN.
1416 * However, if +0 and -0 is handled differently, we cannot use the first one.
1418 if (get_irn_op(cmp) == op_Cmp && get_Cmp_left(cmp) == a) {
1419 if (classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
1420 /* Mux(a CMP 0, X, a) */
1421 if (get_irn_op(b) == op_Minus && get_Minus_op(b) == a) {
1422 /* Mux(a CMP 0, -a, a) */
1423 if (proj_nr == pn_Cmp_Eq) {
1424 /* Mux(a == 0, -a, a) ==> -a */
1426 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1428 else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1429 /* Mux(a != 0, -a, a) ==> a */
1431 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1434 else if (classify_Const(b) == CNST_NULL) {
1435 /* Mux(a CMP 0, 0, a) */
1436 if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1437 /* Mux(a != 0, 0, a) ==> a */
1439 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1441 else if (proj_nr == pn_Cmp_Eq) {
1442 /* Mux(a == 0, 0, a) ==> 0 */
1444 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1454 * Returns a equivalent node of a Psi: if a condition is true
1455 * and all previous conditions are false we know its value.
1456 * If all conditions are false its value is the default one.
1458 static ir_node *equivalent_node_Psi(ir_node *n) {
1460 return equivalent_node_Mux(n);
1465 * Optimize -a CMP -b into b CMP a.
1466 * This works only for for modes where unary Minus
1468 * Note that two-complement integers can Overflow
1469 * so it will NOT work.
1471 static ir_node *equivalent_node_Cmp(ir_node *n)
1473 ir_node *left = get_Cmp_left(n);
1474 ir_node *right = get_Cmp_right(n);
1476 if (get_irn_op(left) == op_Minus && get_irn_op(right) == op_Minus &&
1477 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1478 left = get_Minus_op(left);
1479 right = get_Minus_op(right);
1480 set_Cmp_left(n, right);
1481 set_Cmp_right(n, left);
1487 * Remove Confirm nodes if setting is on.
1488 * Replace Confirms(x, '=', Constlike) by Constlike.
1490 static ir_node *equivalent_node_Confirm(ir_node *n)
1492 ir_node *pred = get_Confirm_value(n);
1493 pn_Cmp pnc = get_Confirm_cmp(n);
1495 if (get_irn_op(pred) == op_Confirm && pnc == get_Confirm_cmp(pred)) {
1497 * rare case: two identical Confirms one after another,
1498 * replace the second one with the first.
1502 if (pnc == pn_Cmp_Eq) {
1503 ir_node *bound = get_Confirm_bound(n);
1506 * Optimize a rare case:
1507 * Confirm(x, '=', Constlike) ==> Constlike
1509 if (is_irn_constlike(bound)) {
1510 DBG_OPT_CONFIRM(n, bound);
1514 return get_opt_remove_confirm() ? get_Confirm_value(n) : n;
1518 * Optimize CopyB(mem, x, x) into a Nop
1520 static ir_node *equivalent_node_CopyB(ir_node *n)
1522 ir_node *a = get_CopyB_dst(n);
1523 ir_node *b = get_CopyB_src(n);
1526 /* Turn CopyB into a tuple (mem, bad, bad) */
1527 ir_node *mem = get_CopyB_mem(n);
1528 turn_into_tuple(n, pn_CopyB_max);
1529 set_Tuple_pred(n, pn_CopyB_M, mem);
1530 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1531 set_Tuple_pred(n, pn_CopyB_M_except, new_Bad());
1537 * Optimize Bounds(idx, idx, upper) into idx.
1539 static ir_node *equivalent_node_Bound(ir_node *n)
1541 ir_node *idx = get_Bound_index(n);
1542 ir_node *lower = get_Bound_lower(n);
1545 /* By definition lower < upper, so if idx == lower -->
1546 lower <= idx && idx < upper */
1548 /* Turn Bound into a tuple (mem, bad, idx) */
1552 ir_node *pred = skip_Proj(idx);
1554 if (get_irn_op(pred) == op_Bound) {
1556 * idx was Bounds_check previously, it is still valid if
1557 * lower <= pred_lower && pred_upper <= upper.
1559 ir_node *upper = get_Bound_upper(n);
1560 if (get_Bound_lower(pred) == lower &&
1561 get_Bound_upper(pred) == upper) {
1563 * One could expect that we simply return the previous
1564 * Bound here. However, this would be wrong, as we could
1565 * add an exception Proj to a new location than.
1566 * So, we must turn in into a tuple
1573 /* Turn Bound into a tuple (mem, bad, idx) */
1574 ir_node *mem = get_Bound_mem(n);
1575 turn_into_tuple(n, pn_Bound_max);
1576 set_Tuple_pred(n, pn_Bound_M, mem);
1577 set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */
1578 set_Tuple_pred(n, pn_Bound_res, idx);
1584 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1585 * perform no actual computation, as, e.g., the Id nodes. It does not create
1586 * new nodes. It is therefore safe to free n if the node returned is not n.
1587 * If a node returns a Tuple we can not just skip it. If the size of the
1588 * in array fits, we transform n into a tuple (e.g., Div).
1591 equivalent_node(ir_node *n)
1593 if (n->op->ops.equivalent_node)
1594 return n->op->ops.equivalent_node(n);
1599 * sets the default equivalent node operation for an ir_op_ops.
1601 * @param code the opcode for the default operation
1602 * @param ops the operations initialized
1607 static ir_op_ops *firm_set_default_equivalent_node(opcode code, ir_op_ops *ops)
1611 ops->equivalent_node = equivalent_node_##a; \
1653 * Do node specific optimizations of nodes predecessors.
1656 optimize_preds(ir_node *n) {
1657 ir_node *a = NULL, *b = NULL;
1659 /* get the operands we will work on for simple cases. */
1661 a = get_binop_left(n);
1662 b = get_binop_right(n);
1663 } else if (is_unop(n)) {
1667 switch (get_irn_opcode(n)) {
1670 /* We don't want Cast as input to Cmp. */
1671 if (get_irn_op(a) == op_Cast) {
1675 if (get_irn_op(b) == op_Cast) {
1677 set_Cmp_right(n, b);
1686 * Returns non-zero if a node is a Phi node
1687 * with all predecessors constant.
1689 static int is_const_Phi(ir_node *n) {
1694 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1695 if (! is_Const(get_irn_n(n, i)))
1701 * Apply an evaluator on a binop with a constant operators (and one Phi).
1703 * @param phi the Phi node
1704 * @param other the other operand
1705 * @param eval an evaluator function
1706 * @param left if non-zero, other is the left operand, else the right
1708 * @return a new Phi node if the conversion was successful, NULL else
1710 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), int left) {
1716 int i, n = get_irn_arity(phi);
1718 NEW_ARR_A(void *, res, n);
1720 for (i = 0; i < n; ++i) {
1721 pred = get_irn_n(phi, i);
1722 tv = get_Const_tarval(pred);
1723 tv = eval(other, tv);
1725 if (tv == tarval_bad) {
1726 /* folding failed, bad */
1733 for (i = 0; i < n; ++i) {
1734 pred = get_irn_n(phi, i);
1735 tv = get_Const_tarval(pred);
1736 tv = eval(tv, other);
1738 if (tv == tarval_bad) {
1739 /* folding failed, bad */
1745 mode = get_irn_mode(phi);
1746 irg = current_ir_graph;
1747 for (i = 0; i < n; ++i) {
1748 pred = get_irn_n(phi, i);
1749 res[i] = new_r_Const_type(irg, get_nodes_block(pred),
1750 mode, res[i], get_Const_type(pred));
1752 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1756 * Apply an evaluator on a unop with a constant operator (a Phi).
1758 * @param phi the Phi node
1759 * @param eval an evaluator function
1761 * @return a new Phi node if the conversion was successful, NULL else
1763 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
1769 int i, n = get_irn_arity(phi);
1771 NEW_ARR_A(void *, res, n);
1772 for (i = 0; i < n; ++i) {
1773 pred = get_irn_n(phi, i);
1774 tv = get_Const_tarval(pred);
1777 if (tv == tarval_bad) {
1778 /* folding failed, bad */
1783 mode = get_irn_mode(phi);
1784 irg = current_ir_graph;
1785 for (i = 0; i < n; ++i) {
1786 pred = get_irn_n(phi, i);
1787 res[i] = new_r_Const_type(irg, get_nodes_block(pred),
1788 mode, res[i], get_Const_type(pred));
1790 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1794 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1795 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1796 * If possible, remove the Conv's.
1798 static ir_node *transform_node_AddSub(ir_node *n)
1800 ir_mode *mode = get_irn_mode(n);
1802 if (mode_is_reference(mode)) {
1803 ir_node *left = get_binop_left(n);
1804 ir_node *right = get_binop_right(n);
1805 int ref_bits = get_mode_size_bits(mode);
1807 if (get_irn_op(left) == op_Conv) {
1808 ir_mode *mode = get_irn_mode(left);
1809 int bits = get_mode_size_bits(mode);
1811 if (ref_bits == bits &&
1812 mode_is_int(mode) &&
1813 get_mode_arithmetic(mode) == irma_twos_complement) {
1814 ir_node *pre = get_Conv_op(left);
1815 ir_mode *pre_mode = get_irn_mode(pre);
1817 if (mode_is_int(pre_mode) &&
1818 get_mode_size_bits(pre_mode) == bits &&
1819 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1820 /* ok, this conv just changes to sign, moreover the calculation
1821 * is done with same number of bits as our address mode, so
1822 * we can ignore the conv as address calculation can be viewed
1823 * as either signed or unsigned
1825 set_binop_left(n, pre);
1830 if (get_irn_op(right) == op_Conv) {
1831 ir_mode *mode = get_irn_mode(right);
1832 int bits = get_mode_size_bits(mode);
1834 if (ref_bits == bits &&
1835 mode_is_int(mode) &&
1836 get_mode_arithmetic(mode) == irma_twos_complement) {
1837 ir_node *pre = get_Conv_op(right);
1838 ir_mode *pre_mode = get_irn_mode(pre);
1840 if (mode_is_int(pre_mode) &&
1841 get_mode_size_bits(pre_mode) == bits &&
1842 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1843 /* ok, this conv just changes to sign, moreover the calculation
1844 * is done with same number of bits as our address mode, so
1845 * we can ignore the conv as address calculation can be viewed
1846 * as either signed or unsigned
1848 set_binop_right(n, pre);
1856 #define HANDLE_BINOP_PHI(op,a,b,c) \
1858 if (is_Const(b) && is_const_Phi(a)) { \
1859 /* check for Op(Phi, Const) */ \
1860 c = apply_binop_on_phi(a, get_Const_tarval(b), op, 0); \
1862 else if (is_Const(a) && is_const_Phi(b)) { \
1863 /* check for Op(Const, Phi) */ \
1864 c = apply_binop_on_phi(b, get_Const_tarval(a), op, 1); \
1867 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1871 #define HANDLE_UNOP_PHI(op,a,c) \
1873 if (is_const_Phi(a)) { \
1874 /* check for Op(Phi) */ \
1875 c = apply_unop_on_phi(a, op); \
1878 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1884 * Do the AddSub optimization, then Transform
1885 * Constant folding on Phi
1886 * Add(a,a) -> Mul(a, 2)
1887 * Add(Mul(a, x), a) -> Mul(a, x+1)
1888 * if the mode is integer or float.
1889 * Transform Add(a,-b) into Sub(a,b).
1890 * Reassociation might fold this further.
1892 static ir_node *transform_node_Add(ir_node *n)
1895 ir_node *a, *b, *c, *oldn = n;
1897 n = transform_node_AddSub(n);
1899 a = get_Add_left(n);
1900 b = get_Add_right(n);
1902 HANDLE_BINOP_PHI(tarval_add, a,b,c);
1904 mode = get_irn_mode(n);
1905 if (mode_is_num(mode)) {
1907 ir_node *block = get_irn_n(n, -1);
1910 get_irn_dbg_info(n),
1914 new_r_Const_long(current_ir_graph, block, mode, 2),
1916 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
1918 else if (get_irn_op(a) == op_Minus) {
1920 get_irn_dbg_info(n),
1926 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1928 else if (get_irn_op(b) == op_Minus) {
1930 get_irn_dbg_info(n),
1936 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1938 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1939 else if (!get_opt_reassociation() && get_irn_op(a) == op_Mul) {
1940 ir_node *ma = get_Mul_left(a);
1941 ir_node *mb = get_Mul_right(a);
1944 ir_node *blk = get_irn_n(n, -1);
1946 get_irn_dbg_info(n), current_ir_graph, blk,
1949 get_irn_dbg_info(n), current_ir_graph, blk,
1951 new_r_Const_long(current_ir_graph, blk, mode, 1),
1954 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1957 ir_node *blk = get_irn_n(n, -1);
1959 get_irn_dbg_info(n), current_ir_graph, blk,
1962 get_irn_dbg_info(n), current_ir_graph, blk,
1964 new_r_Const_long(current_ir_graph, blk, mode, 1),
1967 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1970 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1971 else if (!get_opt_reassociation() && get_irn_op(b) == op_Mul) {
1972 ir_node *ma = get_Mul_left(b);
1973 ir_node *mb = get_Mul_right(b);
1976 ir_node *blk = get_irn_n(n, -1);
1978 get_irn_dbg_info(n), current_ir_graph, blk,
1981 get_irn_dbg_info(n), current_ir_graph, blk,
1983 new_r_Const_long(current_ir_graph, blk, mode, 1),
1986 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1989 ir_node *blk = get_irn_n(n, -1);
1991 get_irn_dbg_info(n), current_ir_graph, blk,
1994 get_irn_dbg_info(n), current_ir_graph, blk,
1996 new_r_Const_long(current_ir_graph, blk, mode, 1),
1999 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
2007 * Do the AddSub optimization, then Transform
2008 * Constant folding on Phi
2009 * Sub(0,a) -> Minus(a)
2010 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2011 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2013 static ir_node *transform_node_Sub(ir_node *n)
2019 n = transform_node_AddSub(n);
2021 a = get_Sub_left(n);
2022 b = get_Sub_right(n);
2024 HANDLE_BINOP_PHI(tarval_sub, a,b,c);
2026 mode = get_irn_mode(n);
2027 if (mode_is_num(mode) && (classify_Const(a) == CNST_NULL)) {
2029 get_irn_dbg_info(n),
2034 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2036 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2037 else if (get_opt_reassociation() && get_irn_op(a) == op_Mul) {
2038 ir_node *ma = get_Mul_left(a);
2039 ir_node *mb = get_Mul_right(a);
2042 ir_node *blk = get_irn_n(n, -1);
2044 get_irn_dbg_info(n),
2045 current_ir_graph, blk,
2048 get_irn_dbg_info(n),
2049 current_ir_graph, blk,
2051 new_r_Const_long(current_ir_graph, blk, mode, 1),
2054 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2057 ir_node *blk = get_irn_n(n, -1);
2059 get_irn_dbg_info(n),
2060 current_ir_graph, blk,
2063 get_irn_dbg_info(n),
2064 current_ir_graph, blk,
2066 new_r_Const_long(current_ir_graph, blk, mode, 1),
2069 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2072 else if (get_irn_op(a) == op_Sub) {
2073 ir_node *x = get_Sub_left(a);
2074 ir_node *y = get_Sub_right(a);
2075 ir_node *blk = get_irn_n(n, -1);
2076 ir_mode *m_b = get_irn_mode(b);
2077 ir_mode *m_y = get_irn_mode(y);
2080 /* Determine the right mode for the Add. */
2083 else if (mode_is_reference(m_b))
2085 else if (mode_is_reference(m_y))
2089 * Both modes are different but none is reference,
2090 * happens for instance in SubP(SubP(P, Iu), Is).
2091 * We have two possibilities here: Cast or ignore.
2092 * Currently we ignore this case.
2097 add = new_r_Add(current_ir_graph, blk, y, b, mode);
2100 set_Sub_right(n, add);
2101 DBG_OPT_ALGSIM0(n, n, FS_OPT_SUB_SUB_X_Y_Z);
2108 * Transform Mul(a,-1) into -a.
2109 * Do constant evaluation of Phi nodes.
2110 * Do architecture dependent optimizations on Mul nodes
2112 static ir_node *transform_node_Mul(ir_node *n) {
2113 ir_node *c, *oldn = n;
2114 ir_node *a = get_Mul_left(n);
2115 ir_node *b = get_Mul_right(n);
2118 HANDLE_BINOP_PHI(tarval_mul, a,b,c);
2120 mode = get_irn_mode(n);
2121 if (mode_is_signed(mode)) {
2124 if (value_of(a) == get_mode_minus_one(mode))
2126 else if (value_of(b) == get_mode_minus_one(mode))
2129 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2130 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2134 return arch_dep_replace_mul_with_shifts(n);
2138 * Transform a Div Node.
2140 static ir_node *transform_node_Div(ir_node *n)
2142 tarval *tv = value_of(n);
2145 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2147 if (tv != tarval_bad) {
2148 value = new_Const(get_tarval_mode(tv), tv);
2150 DBG_OPT_CSTEVAL(n, value);
2152 else /* Try architecture dependent optimization */
2153 value = arch_dep_replace_div_by_const(n);
2156 /* Turn Div into a tuple (mem, bad, value) */
2157 ir_node *mem = get_Div_mem(n);
2159 turn_into_tuple(n, pn_Div_max);
2160 set_Tuple_pred(n, pn_Div_M, mem);
2161 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2162 set_Tuple_pred(n, pn_Div_res, value);
2168 * Transform a Mod node.
2170 static ir_node *transform_node_Mod(ir_node *n)
2172 tarval *tv = value_of(n);
2175 /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */
2177 if (tv != tarval_bad) {
2178 value = new_Const(get_tarval_mode(tv), tv);
2180 DBG_OPT_CSTEVAL(n, value);
2182 else /* Try architecture dependent optimization */
2183 value = arch_dep_replace_mod_by_const(n);
2186 /* Turn Mod into a tuple (mem, bad, value) */
2187 ir_node *mem = get_Mod_mem(n);
2189 turn_into_tuple(n, pn_Mod_max);
2190 set_Tuple_pred(n, pn_Mod_M, mem);
2191 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2192 set_Tuple_pred(n, pn_Mod_res, value);
2198 * Transform a DivMod node.
2200 static ir_node *transform_node_DivMod(ir_node *n)
2204 ir_node *a = get_DivMod_left(n);
2205 ir_node *b = get_DivMod_right(n);
2206 ir_mode *mode = get_irn_mode(a);
2207 tarval *ta = value_of(a);
2208 tarval *tb = value_of(b);
2210 if (!(mode_is_int(mode) && mode_is_int(get_irn_mode(b))))
2213 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2215 if (tb != tarval_bad) {
2216 if (tb == get_mode_one(get_tarval_mode(tb))) {
2217 b = new_Const (mode, get_mode_null(mode));
2220 DBG_OPT_CSTEVAL(n, b);
2222 else if (ta != tarval_bad) {
2223 tarval *resa, *resb;
2224 resa = tarval_div (ta, tb);
2225 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2226 Jmp for X result!? */
2227 resb = tarval_mod (ta, tb);
2228 if (resb == tarval_bad) return n; /* Causes exception! */
2229 a = new_Const (mode, resa);
2230 b = new_Const (mode, resb);
2233 DBG_OPT_CSTEVAL(n, a);
2234 DBG_OPT_CSTEVAL(n, b);
2236 else { /* Try architecture dependent optimization */
2237 arch_dep_replace_divmod_by_const(&a, &b, n);
2238 evaluated = a != NULL;
2240 } else if (ta == get_mode_null(mode)) {
2241 /* 0 / non-Const = 0 */
2246 if (evaluated) { /* replace by tuple */
2247 ir_node *mem = get_DivMod_mem(n);
2248 turn_into_tuple(n, pn_DivMod_max);
2249 set_Tuple_pred(n, pn_DivMod_M, mem);
2250 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
2251 set_Tuple_pred(n, pn_DivMod_res_div, a);
2252 set_Tuple_pred(n, pn_DivMod_res_mod, b);
2259 * Optimize Abs(x) into x if x is Confirmed >= 0
2260 * Optimize Abs(x) into -x if x is Confirmed <= 0
2262 static ir_node *transform_node_Abs(ir_node *n)
2265 ir_node *a = get_Abs_op(n);
2266 value_classify sign = classify_value_sign(a);
2268 if (sign == VALUE_NEGATIVE) {
2269 ir_mode *mode = get_irn_mode(n);
2272 * We can replace the Abs by -x here.
2273 * We even could add a new Confirm here.
2275 * Note that -x would create a new node, so we could
2276 * not run it in the equivalent_node() context.
2278 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
2279 get_irn_n(n, -1), a, mode);
2281 DBG_OPT_CONFIRM(oldn, n);
2283 else if (sign == VALUE_POSITIVE) {
2284 /* n is positive, Abs is not needed */
2287 DBG_OPT_CONFIRM(oldn, n);
2294 * Transform a Cond node.
2296 static ir_node *transform_node_Cond(ir_node *n)
2298 /* Replace the Cond by a Jmp if it branches on a constant
2301 ir_node *a = get_Cond_selector(n);
2302 tarval *ta = value_of(a);
2304 /* we need block info which is not available in floating irgs */
2305 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
2308 if ((ta != tarval_bad) &&
2309 (get_irn_mode(a) == mode_b) &&
2310 (get_opt_unreachable_code())) {
2311 /* It's a boolean Cond, branching on a boolean constant.
2312 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2313 jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n));
2314 turn_into_tuple(n, pn_Cond_max);
2315 if (ta == tarval_b_true) {
2316 set_Tuple_pred(n, pn_Cond_false, new_Bad());
2317 set_Tuple_pred(n, pn_Cond_true, jmp);
2319 set_Tuple_pred(n, pn_Cond_false, jmp);
2320 set_Tuple_pred(n, pn_Cond_true, new_Bad());
2322 /* We might generate an endless loop, so keep it alive. */
2323 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
2331 static ir_node *transform_node_And(ir_node *n)
2333 ir_node *c, *oldn = n;
2334 ir_node *a = get_And_left(n);
2335 ir_node *b = get_And_right(n);
2337 HANDLE_BINOP_PHI(tarval_and, a,b,c);
2344 static ir_node *transform_node_Eor(ir_node *n)
2346 ir_node *c, *oldn = n;
2347 ir_node *a = get_Eor_left(n);
2348 ir_node *b = get_Eor_right(n);
2349 ir_mode *mode = get_irn_mode(n);
2351 HANDLE_BINOP_PHI(tarval_eor, a,b,c);
2355 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
2356 mode, get_mode_null(mode));
2357 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
2359 else if ((mode == mode_b)
2360 && (get_irn_op(a) == op_Proj)
2361 && (get_irn_mode(a) == mode_b)
2362 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)
2363 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2364 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
2365 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2366 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
2368 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
2370 else if ((mode == mode_b)
2371 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) {
2372 /* The Eor is a Not. Replace it by a Not. */
2373 /* ????!!!Extend to bitfield 1111111. */
2374 n = new_r_Not(current_ir_graph, get_irn_n(n, -1), a, mode_b);
2376 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2385 static ir_node *transform_node_Not(ir_node *n)
2387 ir_node *c, *oldn = n;
2388 ir_node *a = get_Not_op(n);
2390 HANDLE_UNOP_PHI(tarval_not,a,c);
2392 /* check for a boolean Not */
2393 if ( (get_irn_mode(n) == mode_b)
2394 && (get_irn_op(a) == op_Proj)
2395 && (get_irn_mode(a) == mode_b)
2396 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2397 /* We negate a Cmp. The Cmp has the negated result anyways! */
2398 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2399 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
2400 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
2406 * Transform a Minus.
2408 static ir_node *transform_node_Minus(ir_node *n)
2410 ir_node *c, *oldn = n;
2411 ir_node *a = get_Minus_op(n);
2413 HANDLE_UNOP_PHI(tarval_neg,a,c);
2418 * Transform a Cast_type(Const) into a new Const_type
2420 static ir_node *transform_node_Cast(ir_node *n) {
2422 ir_node *pred = get_Cast_op(n);
2423 ir_type *tp = get_irn_type(n);
2425 if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) {
2426 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
2427 get_Const_tarval(pred), tp);
2428 DBG_OPT_CSTEVAL(oldn, n);
2429 } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) {
2430 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred),
2431 get_SymConst_kind(pred), tp);
2432 DBG_OPT_CSTEVAL(oldn, n);
2439 * Transform a Proj(Div) with a non-zero value.
2440 * Removes the exceptions and routes the memory to the NoMem node.
2442 static ir_node *transform_node_Proj_Div(ir_node *proj)
2444 ir_node *n = get_Proj_pred(proj);
2445 ir_node *b = get_Div_right(n);
2448 if (value_not_zero(b)) {
2449 /* div(x, y) && y != 0 */
2450 proj_nr = get_Proj_proj(proj);
2452 /* this node may float if it did not depend on a Confirm */
2453 set_irn_pinned(n, op_pin_state_floats);
2455 if (proj_nr == pn_Div_X_except) {
2456 /* we found an exception handler, remove it */
2457 DBG_OPT_EXC_REM(proj);
2460 else if (proj_nr == pn_Div_M) {
2461 ir_node *res = get_Div_mem(n);
2462 /* the memory Proj can only be removed if we divide by a
2463 real constant, but the node never produce a new memory */
2464 if (value_of(b) != tarval_bad) {
2465 /* this is a Div by a const, we can remove the memory edge */
2466 set_Div_mem(n, get_irg_no_mem(current_ir_graph));
2475 * Transform a Proj(Mod) with a non-zero value.
2476 * Removes the exceptions and routes the memory to the NoMem node.
2478 static ir_node *transform_node_Proj_Mod(ir_node *proj)
2480 ir_node *n = get_Proj_pred(proj);
2481 ir_node *b = get_Mod_right(n);
2484 if (value_not_zero(b)) {
2485 /* mod(x, y) && y != 0 */
2486 proj_nr = get_Proj_proj(proj);
2488 /* this node may float if it did not depend on a Confirm */
2489 set_irn_pinned(n, op_pin_state_floats);
2491 if (proj_nr == pn_Mod_X_except) {
2492 /* we found an exception handler, remove it */
2493 DBG_OPT_EXC_REM(proj);
2495 } else if (proj_nr == pn_Mod_M) {
2496 ir_node *res = get_Mod_mem(n);
2497 /* the memory Proj can only be removed if we divide by a
2498 real constant, but the node never produce a new memory */
2499 if (value_of(b) != tarval_bad) {
2500 /* this is a Mod by a const, we can remove the memory edge */
2501 set_Mod_mem(n, get_irg_no_mem(current_ir_graph));
2505 else if (proj_nr == pn_Mod_res && get_Mod_left(n) == b) {
2506 /* a % a = 0 if a != 0 */
2507 ir_mode *mode = get_irn_mode(proj);
2508 ir_node *res = new_Const(mode, get_mode_null(mode));
2510 DBG_OPT_CSTEVAL(n, res);
2518 * Transform a Proj(DivMod) with a non-zero value.
2519 * Removes the exceptions and routes the memory to the NoMem node.
2521 static ir_node *transform_node_Proj_DivMod(ir_node *proj)
2523 ir_node *n = get_Proj_pred(proj);
2524 ir_node *b = get_DivMod_right(n);
2527 if (value_not_zero(b)) {
2528 /* DivMod(x, y) && y != 0 */
2529 proj_nr = get_Proj_proj(proj);
2531 /* this node may float if it did not depend on a Confirm */
2532 set_irn_pinned(n, op_pin_state_floats);
2534 if (proj_nr == pn_DivMod_X_except) {
2535 /* we found an exception handler, remove it */
2536 DBG_OPT_EXC_REM(proj);
2539 else if (proj_nr == pn_DivMod_M) {
2540 ir_node *res = get_DivMod_mem(n);
2541 /* the memory Proj can only be removed if we divide by a
2542 real constant, but the node never produce a new memory */
2543 if (value_of(b) != tarval_bad) {
2544 /* this is a DivMod by a const, we can remove the memory edge */
2545 set_DivMod_mem(n, get_irg_no_mem(current_ir_graph));
2549 else if (proj_nr == pn_DivMod_res_mod && get_DivMod_left(n) == b) {
2550 /* a % a = 0 if a != 0 */
2551 ir_mode *mode = get_irn_mode(proj);
2552 ir_node *res = new_Const(mode, get_mode_null(mode));
2554 DBG_OPT_CSTEVAL(n, res);
2562 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
2564 static ir_node *transform_node_Proj_Cond(ir_node *proj)
2566 if (get_opt_unreachable_code()) {
2567 ir_node *n = get_Proj_pred(proj);
2568 ir_node *b = get_Cond_selector(n);
2570 if (mode_is_int(get_irn_mode(b))) {
2571 tarval *tb = value_of(b);
2573 if (tb != tarval_bad) {
2574 /* we have a constant switch */
2575 long num = get_Proj_proj(proj);
2577 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
2578 if (get_tarval_long(tb) == num) {
2579 /* Do NOT create a jump here, or we will have 2 control flow ops
2580 * in a block. This case is optimized away in optimize_cf(). */
2584 /* this case will NEVER be taken, kill it */
2595 * Normalizes and optimizes Cmp nodes.
2597 static ir_node *transform_node_Proj_Cmp(ir_node *proj)
2599 if (get_opt_reassociation()) {
2600 ir_node *n = get_Proj_pred(proj);
2601 ir_node *left = get_Cmp_left(n);
2602 ir_node *right = get_Cmp_right(n);
2606 ir_mode *mode = NULL;
2607 long proj_nr = get_Proj_proj(proj);
2610 * First step: normalize the compare op
2611 * by placing the constant on the right site
2612 * or moving the lower address node to the left.
2613 * We ignore the case that both are constants
2614 * this case should be optimized away.
2616 if (get_irn_op(right) == op_Const)
2618 else if (get_irn_op(left) == op_Const) {
2623 proj_nr = get_inversed_pnc(proj_nr);
2626 else if (get_irn_idx(left) > get_irn_idx(right)) {
2632 proj_nr = get_inversed_pnc(proj_nr);
2637 * Second step: Try to reduce the magnitude
2638 * of a constant. This may help to generate better code
2639 * later and may help to normalize more compares.
2640 * Of course this is only possible for integer values.
2643 mode = get_irn_mode(c);
2644 tv = get_Const_tarval(c);
2646 if (tv != tarval_bad) {
2647 /* the following optimization is possible on modes without Overflow
2648 * on Unary Minus or on == and !=:
2649 * -a CMP c ==> a swap(CMP) -c
2651 * Beware: for two-complement Overflow may occur, so only == and != can
2652 * be optimized, see this:
2653 * -MININT < 0 =/=> MININT > 0 !!!
2655 if (get_opt_constant_folding() && get_irn_op(left) == op_Minus &&
2656 (!mode_overflow_on_unary_Minus(mode) ||
2657 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
2658 left = get_Minus_op(left);
2659 tv = tarval_sub(get_mode_null(mode), tv);
2661 proj_nr = get_inversed_pnc(proj_nr);
2665 /* for integer modes, we have more */
2666 if (mode_is_int(mode)) {
2667 /* Ne includes Unordered which is not possible on integers.
2668 * However, frontends often use this wrong, so fix it here */
2669 if (proj_nr & pn_Cmp_Uo) {
2670 proj_nr &= ~pn_Cmp_Uo;
2671 set_Proj_proj(proj, proj_nr);
2674 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
2675 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
2676 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
2677 tv = tarval_sub(tv, get_mode_one(mode));
2679 proj_nr ^= pn_Cmp_Eq;
2682 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
2683 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
2684 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
2685 tv = tarval_add(tv, get_mode_one(mode));
2687 proj_nr ^= pn_Cmp_Eq;
2691 /* the following reassociations work only for == and != */
2692 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
2694 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
2695 if (classify_tarval(tv) == TV_CLASSIFY_NULL && get_irn_op(left) == op_Sub) {
2696 right = get_Sub_right(left);
2697 left = get_Sub_left(left);
2699 tv = value_of(right);
2703 if (tv != tarval_bad) {
2704 ir_op *op = get_irn_op(left);
2706 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
2708 ir_node *c1 = get_Sub_right(left);
2709 tarval *tv2 = value_of(c1);
2711 if (tv2 != tarval_bad) {
2712 tv2 = tarval_add(tv, value_of(c1));
2714 if (tv2 != tarval_bad) {
2715 left = get_Sub_left(left);
2721 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
2722 else if (op == op_Add) {
2723 ir_node *a_l = get_Add_left(left);
2724 ir_node *a_r = get_Add_right(left);
2728 if (get_irn_op(a_l) == op_Const) {
2730 tv2 = value_of(a_l);
2734 tv2 = value_of(a_r);
2737 if (tv2 != tarval_bad) {
2738 tv2 = tarval_sub(tv, tv2);
2740 if (tv2 != tarval_bad) {
2747 /* -a == c ==> a == -c, -a != c ==> a != -c */
2748 else if (op == op_Minus) {
2749 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
2751 if (tv2 != tarval_bad) {
2752 left = get_Minus_op(left);
2759 /* the following reassociations work only for <= */
2760 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2761 if (tv != tarval_bad) {
2762 ir_op *op = get_irn_op(left);
2764 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
2772 * optimization for AND:
2774 * And(x, C) == C ==> And(x, C) != 0
2775 * And(x, C) != C ==> And(X, C) == 0
2777 * if C is a single Bit constant.
2779 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) &&
2780 (get_irn_op(left) == op_And)) {
2781 if (is_single_bit_tarval(tv)) {
2782 /* check for Constant's match. We have check hare the tarvals,
2783 because our const might be changed */
2784 ir_node *la = get_And_left(left);
2785 ir_node *ra = get_And_right(left);
2786 if ((is_Const(la) && get_Const_tarval(la) == tv) ||
2787 (is_Const(ra) && get_Const_tarval(ra) == tv)) {
2788 /* fine: do the transformation */
2789 tv = get_mode_null(get_tarval_mode(tv));
2790 proj_nr ^= pn_Cmp_Leg;
2795 } /* tarval != bad */
2799 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
2801 if (changed & 2) /* need a new Const */
2802 right = new_Const(mode, tv);
2804 /* create a new compare */
2805 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block,
2808 set_Proj_pred(proj, n);
2809 set_Proj_proj(proj, proj_nr);
2816 * Does all optimizations on nodes that must be done on it's Proj's
2817 * because of creating new nodes.
2819 static ir_node *transform_node_Proj(ir_node *proj)
2821 ir_node *n = get_Proj_pred(proj);
2823 switch (get_irn_opcode(n)) {
2825 return transform_node_Proj_Div(proj);
2828 return transform_node_Proj_Mod(proj);
2831 return transform_node_Proj_DivMod(proj);
2834 return transform_node_Proj_Cond(proj);
2837 return transform_node_Proj_Cmp(proj);
2840 /* should not happen, but if it does will be optimized away */
2841 return equivalent_node_Proj(proj);
2850 * Move Confirms down through Phi nodes.
2852 static ir_node *transform_node_Phi(ir_node *phi) {
2854 ir_mode *mode = get_irn_mode(phi);
2856 if (mode_is_reference(mode)) {
2857 n = get_irn_arity(phi);
2859 /* Beware of Phi0 */
2861 ir_node *pred = get_irn_n(phi, 0);
2862 ir_node *bound, *new_Phi, *block, **in;
2865 if (! is_Confirm(pred))
2868 bound = get_Confirm_bound(pred);
2869 pnc = get_Confirm_cmp(pred);
2871 NEW_ARR_A(ir_node *, in, n);
2872 in[0] = get_Confirm_value(pred);
2874 for (i = 1; i < n; ++i) {
2875 pred = get_irn_n(phi, i);
2877 if (! is_Confirm(pred) ||
2878 get_Confirm_bound(pred) != bound ||
2879 get_Confirm_cmp(pred) != pnc)
2881 in[i] = get_Confirm_value(pred);
2883 /* move the Confirm nodes "behind" the Phi */
2884 block = get_irn_n(phi, -1);
2885 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
2886 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
2893 * returns the operands of a commutative bin-op, if one operand is
2894 * a const, it is returned as the second one.
2896 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c)
2898 ir_node *op_a = get_binop_left(binop);
2899 ir_node *op_b = get_binop_right(binop);
2901 assert(is_op_commutative(get_irn_op(binop)));
2903 if (get_irn_op(op_a) == op_Const) {
2914 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
2915 * Such pattern may arise in bitfield stores.
2917 * value c4 value c4 & c2
2918 * AND c3 AND c1 | c3
2923 static ir_node *transform_node_Or_bf_store(ir_node *or)
2927 ir_node *and_l, *c3;
2928 ir_node *value, *c4;
2929 ir_node *new_and, *new_const, *block;
2930 ir_mode *mode = get_irn_mode(or);
2932 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
2934 get_comm_Binop_Ops(or, &and, &c1);
2935 if ((get_irn_op(c1) != op_Const) || (get_irn_op(and) != op_And))
2938 get_comm_Binop_Ops(and, &or_l, &c2);
2939 if ((get_irn_op(c2) != op_Const) || (get_irn_op(or_l) != op_Or))
2942 get_comm_Binop_Ops(or_l, &and_l, &c3);
2943 if ((get_irn_op(c3) != op_Const) || (get_irn_op(and_l) != op_And))
2946 get_comm_Binop_Ops(and_l, &value, &c4);
2947 if (get_irn_op(c4) != op_Const)
2950 /* ok, found the pattern, check for conditions */
2951 assert(mode == get_irn_mode(and));
2952 assert(mode == get_irn_mode(or_l));
2953 assert(mode == get_irn_mode(and_l));
2955 tv1 = get_Const_tarval(c1);
2956 tv2 = get_Const_tarval(c2);
2957 tv3 = get_Const_tarval(c3);
2958 tv4 = get_Const_tarval(c4);
2960 tv = tarval_or(tv4, tv2);
2961 if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) {
2962 /* have at least one 0 at the same bit position */
2966 n_tv4 = tarval_not(tv4);
2967 if (tv3 != tarval_and(tv3, n_tv4)) {
2968 /* bit in the or_mask is outside the and_mask */
2972 n_tv2 = tarval_not(tv2);
2973 if (tv1 != tarval_and(tv1, n_tv2)) {
2974 /* bit in the or_mask is outside the and_mask */
2978 /* ok, all conditions met */
2979 block = get_irn_n(or, -1);
2981 new_and = new_r_And(current_ir_graph, block,
2982 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
2984 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
2986 set_Or_left(or, new_and);
2987 set_Or_right(or, new_const);
2989 /* check for more */
2990 return transform_node_Or_bf_store(or);
2994 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
2996 static ir_node *transform_node_Or_Rot(ir_node *or)
2998 ir_mode *mode = get_irn_mode(or);
2999 ir_node *shl, *shr, *block;
3000 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
3003 if (! mode_is_int(mode))
3006 shl = get_binop_left(or);
3007 shr = get_binop_right(or);
3009 if (get_irn_op(shl) == op_Shr) {
3010 if (get_irn_op(shr) != op_Shl)
3017 else if (get_irn_op(shl) != op_Shl)
3019 else if (get_irn_op(shr) != op_Shr)
3022 x = get_Shl_left(shl);
3023 if (x != get_Shr_left(shr))
3026 c1 = get_Shl_right(shl);
3027 c2 = get_Shr_right(shr);
3028 if (get_irn_op(c1) == op_Const && get_irn_op(c2) == op_Const) {
3029 tv1 = get_Const_tarval(c1);
3030 if (! tarval_is_long(tv1))
3033 tv2 = get_Const_tarval(c2);
3034 if (! tarval_is_long(tv2))
3037 if (get_tarval_long(tv1) + get_tarval_long(tv2)
3038 != get_mode_size_bits(mode))
3041 /* yet, condition met */
3042 block = get_irn_n(or, -1);
3044 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
3046 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
3049 else if (get_irn_op(c1) == op_Sub) {
3053 if (get_Sub_right(sub) != v)
3056 c1 = get_Sub_left(sub);
3057 if (get_irn_op(c1) != op_Const)
3060 tv1 = get_Const_tarval(c1);
3061 if (! tarval_is_long(tv1))
3064 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
3067 /* yet, condition met */
3068 block = get_nodes_block(or);
3070 /* a Rot right is not supported, so use a rot left */
3071 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
3073 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
3076 else if (get_irn_op(c2) == op_Sub) {
3080 c1 = get_Sub_left(sub);
3081 if (get_irn_op(c1) != op_Const)
3084 tv1 = get_Const_tarval(c1);
3085 if (! tarval_is_long(tv1))
3088 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
3091 /* yet, condition met */
3092 block = get_irn_n(or, -1);
3095 n = new_r_Rot(current_ir_graph, block, x, v, mode);
3097 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
3107 static ir_node *transform_node_Or(ir_node *n)
3109 ir_node *c, *oldn = n;
3110 ir_node *a = get_Or_left(n);
3111 ir_node *b = get_Or_right(n);
3113 HANDLE_BINOP_PHI(tarval_or, a,b,c);
3115 n = transform_node_Or_bf_store(n);
3116 n = transform_node_Or_Rot(n);
3123 static ir_node *transform_node(ir_node *n);
3126 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl.
3128 * Should be moved to reassociation?
3130 static ir_node *transform_node_shift(ir_node *n)
3132 ir_node *left, *right;
3133 tarval *tv1, *tv2, *res;
3135 int modulo_shf, flag;
3137 left = get_binop_left(n);
3139 /* different operations */
3140 if (get_irn_op(left) != get_irn_op(n))
3143 right = get_binop_right(n);
3144 tv1 = value_of(right);
3145 if (tv1 == tarval_bad)
3148 tv2 = value_of(get_binop_right(left));
3149 if (tv2 == tarval_bad)
3152 res = tarval_add(tv1, tv2);
3154 /* beware: a simple replacement works only, if res < modulo shift */
3155 mode = get_irn_mode(n);
3159 modulo_shf = get_mode_modulo_shift(mode);
3160 if (modulo_shf > 0) {
3161 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
3163 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
3170 /* ok, we can replace it */
3171 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
3173 in[0] = get_binop_left(left);
3174 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
3176 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
3178 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
3180 return transform_node(irn);
3188 static ir_node *transform_node_Shr(ir_node *n)
3190 ir_node *c, *oldn = n;
3191 ir_node *a = get_Shr_left(n);
3192 ir_node *b = get_Shr_right(n);
3194 HANDLE_BINOP_PHI(tarval_shr, a, b, c);
3195 return transform_node_shift(n);
3201 static ir_node *transform_node_Shrs(ir_node *n)
3203 ir_node *c, *oldn = n;
3204 ir_node *a = get_Shrs_left(n);
3205 ir_node *b = get_Shrs_right(n);
3207 HANDLE_BINOP_PHI(tarval_shrs, a, b, c);
3208 return transform_node_shift(n);
3214 static ir_node *transform_node_Shl(ir_node *n)
3216 ir_node *c, *oldn = n;
3217 ir_node *a = get_Shl_left(n);
3218 ir_node *b = get_Shl_right(n);
3220 HANDLE_BINOP_PHI(tarval_shl, a, b, c);
3221 return transform_node_shift(n);
3225 * Remove dead blocks and nodes in dead blocks
3226 * in keep alive list. We do not generate a new End node.
3228 static ir_node *transform_node_End(ir_node *n) {
3229 int i, n_keepalives = get_End_n_keepalives(n);
3231 for (i = 0; i < n_keepalives; ++i) {
3232 ir_node *ka = get_End_keepalive(n, i);
3234 if (is_Block_dead(ka)) {
3235 set_End_keepalive(n, i, new_Bad());
3238 else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka)))
3239 set_End_keepalive(n, i, new_Bad());
3245 * Optimize a Mux into some simpler cases.
3247 static ir_node *transform_node_Mux(ir_node *n)
3249 ir_node *oldn = n, *sel = get_Mux_sel(n);
3250 ir_mode *mode = get_irn_mode(n);
3252 if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(mode)) {
3253 ir_node *cmp = get_Proj_pred(sel);
3254 long proj_nr = get_Proj_proj(sel);
3255 ir_node *f = get_Mux_false(n);
3256 ir_node *t = get_Mux_true(n);
3258 if (get_irn_op(cmp) == op_Cmp && classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
3259 ir_node *block = get_irn_n(n, -1);
3262 * Note: normalization puts the constant on the right site,
3263 * so we check only one case.
3265 * Note further that these optimization work even for floating point
3266 * with NaN's because -NaN == NaN.
3267 * However, if +0 and -0 is handled differently, we cannot use the first one.
3269 if (get_irn_op(f) == op_Minus &&
3270 get_Minus_op(f) == t &&
3271 get_Cmp_left(cmp) == t) {
3273 if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3274 /* Mux(a >=/> 0, -a, a) ==> Abs(a) */
3275 n = new_rd_Abs(get_irn_dbg_info(n),
3279 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3282 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3283 /* Mux(a <=/< 0, -a, a) ==> Minus(Abs(a)) */
3284 n = new_rd_Abs(get_irn_dbg_info(n),
3288 n = new_rd_Minus(get_irn_dbg_info(n),
3293 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3297 else if (get_irn_op(t) == op_Minus &&
3298 get_Minus_op(t) == f &&
3299 get_Cmp_left(cmp) == f) {
3301 if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3302 /* Mux(a <=/< 0, a, -a) ==> Abs(a) */
3303 n = new_rd_Abs(get_irn_dbg_info(n),
3307 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3310 else if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3311 /* Mux(a >=/> 0, a, -a) ==> Minus(Abs(a)) */
3312 n = new_rd_Abs(get_irn_dbg_info(n),
3316 n = new_rd_Minus(get_irn_dbg_info(n),
3321 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3326 if (mode_is_int(mode) && mode_is_signed(mode) &&
3327 get_mode_arithmetic(mode) == irma_twos_complement) {
3328 ir_node *x = get_Cmp_left(cmp);
3330 /* the following optimization works only with signed integer two-complement mode */
3332 if (mode == get_irn_mode(x)) {
3334 * FIXME: this restriction is two rigid, as it would still
3335 * work if mode(x) = Hs and mode == Is, but at least it removes
3338 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Le) &&
3339 classify_Const(t) == CNST_ALL_ONE &&
3340 classify_Const(f) == CNST_NULL) {
3342 * Mux(x:T </<= 0, 0, -1) -> Shrs(x, sizeof_bits(T) - 1)
3346 n = new_rd_Shrs(get_irn_dbg_info(n),
3347 current_ir_graph, block, x,
3348 new_r_Const_long(current_ir_graph, block, mode_Iu,
3349 get_mode_size_bits(mode) - 1),
3351 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3354 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) &&
3355 classify_Const(t) == CNST_ONE &&
3356 classify_Const(f) == CNST_NULL) {
3358 * Mux(x:T >/>= 0, 0, 1) -> Shr(-x, sizeof_bits(T) - 1)
3362 n = new_rd_Shr(get_irn_dbg_info(n),
3363 current_ir_graph, block,
3364 new_r_Minus(current_ir_graph, block, x, mode),
3365 new_r_Const_long(current_ir_graph, block, mode_Iu,
3366 get_mode_size_bits(mode) - 1),
3368 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3375 return arch_transform_node_Mux(n);
3379 * Optimize a Psi into some simpler cases.
3381 static ir_node *transform_node_Psi(ir_node *n) {
3383 return transform_node_Mux(n);
3389 * Tries several [inplace] [optimizing] transformations and returns an
3390 * equivalent node. The difference to equivalent_node() is that these
3391 * transformations _do_ generate new nodes, and thus the old node must
3392 * not be freed even if the equivalent node isn't the old one.
3394 static ir_node *transform_node(ir_node *n)
3396 if (n->op->ops.transform_node)
3397 n = n->op->ops.transform_node(n);
3402 * sSets the default transform node operation for an ir_op_ops.
3404 * @param code the opcode for the default operation
3405 * @param ops the operations initialized
3410 static ir_op_ops *firm_set_default_transform_node(opcode code, ir_op_ops *ops)
3414 ops->transform_node = transform_node_##a; \
3450 /* **************** Common Subexpression Elimination **************** */
3452 /** The size of the hash table used, should estimate the number of nodes
3454 #define N_IR_NODES 512
3456 /** Compares the attributes of two Const nodes. */
3457 static int node_cmp_attr_Const(ir_node *a, ir_node *b)
3459 return (get_Const_tarval(a) != get_Const_tarval(b))
3460 || (get_Const_type(a) != get_Const_type(b));
3463 /** Compares the attributes of two Proj nodes. */
3464 static int node_cmp_attr_Proj(ir_node *a, ir_node *b)
3466 return get_irn_proj_attr (a) != get_irn_proj_attr (b);
3469 /** Compares the attributes of two Filter nodes. */
3470 static int node_cmp_attr_Filter(ir_node *a, ir_node *b)
3472 return get_Filter_proj(a) != get_Filter_proj(b);
3475 /** Compares the attributes of two Alloc nodes. */
3476 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b)
3478 return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where)
3479 || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type);
3482 /** Compares the attributes of two Free nodes. */
3483 static int node_cmp_attr_Free(ir_node *a, ir_node *b)
3485 return (get_irn_free_attr(a).where != get_irn_free_attr(b).where)
3486 || (get_irn_free_attr(a).type != get_irn_free_attr(b).type);
3489 /** Compares the attributes of two SymConst nodes. */
3490 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b)
3492 return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num)
3493 || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p)
3494 || (get_irn_symconst_attr(a).tp != get_irn_symconst_attr(b).tp);
3497 /** Compares the attributes of two Call nodes. */
3498 static int node_cmp_attr_Call(ir_node *a, ir_node *b)
3500 return (get_irn_call_attr(a) != get_irn_call_attr(b));
3503 /** Compares the attributes of two Sel nodes. */
3504 static int node_cmp_attr_Sel(ir_node *a, ir_node *b)
3506 return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind)
3507 || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name)
3508 || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner)
3509 || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name)
3510 || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type);
3513 /** Compares the attributes of two Phi nodes. */
3514 static int node_cmp_attr_Phi(ir_node *a, ir_node *b)
3516 return get_irn_phi_attr (a) != get_irn_phi_attr (b);
3519 /** Compares the attributes of two Cast nodes. */
3520 static int node_cmp_attr_Cast(ir_node *a, ir_node *b)
3522 return get_Cast_type(a) != get_Cast_type(b);
3525 /** Compares the attributes of two Load nodes. */
3526 static int node_cmp_attr_Load(ir_node *a, ir_node *b)
3528 if (get_Load_volatility(a) == volatility_is_volatile ||
3529 get_Load_volatility(b) == volatility_is_volatile)
3530 /* NEVER do CSE on volatile Loads */
3533 return get_Load_mode(a) != get_Load_mode(b);
3536 /** Compares the attributes of two Store nodes. */
3537 static int node_cmp_attr_Store(ir_node *a, ir_node *b)
3539 /* NEVER do CSE on volatile Stores */
3540 return (get_Store_volatility(a) == volatility_is_volatile ||
3541 get_Store_volatility(b) == volatility_is_volatile);
3544 /** Compares the attributes of two Confirm nodes. */
3545 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b)
3547 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
3551 * Set the default node attribute compare operation for an ir_op_ops.
3553 * @param code the opcode for the default operation
3554 * @param ops the operations initialized
3559 static ir_op_ops *firm_set_default_node_cmp_attr(opcode code, ir_op_ops *ops)
3563 ops->node_cmp_attr = node_cmp_attr_##a; \
3589 * Compare function for two nodes in the hash table. Gets two
3590 * nodes as parameters. Returns 0 if the nodes are a cse.
3592 int identities_cmp(const void *elt, const void *key)
3600 if (a == b) return 0;
3602 if ((get_irn_op(a) != get_irn_op(b)) ||
3603 (get_irn_mode(a) != get_irn_mode(b))) return 1;
3605 /* compare if a's in and b's in are of equal length */
3606 irn_arity_a = get_irn_intra_arity (a);
3607 if (irn_arity_a != get_irn_intra_arity(b))
3610 /* for block-local cse and op_pin_state_pinned nodes: */
3611 if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) {
3612 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
3616 /* compare a->in[0..ins] with b->in[0..ins] */
3617 for (i = 0; i < irn_arity_a; i++)
3618 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
3622 * here, we already now that the nodes are identical except their
3625 if (a->op->ops.node_cmp_attr)
3626 return a->op->ops.node_cmp_attr(a, b);
3632 * Calculate a hash value of a node.
3634 unsigned ir_node_hash(ir_node *node)
3639 if (node->op == op_Const) {
3640 /* special value for const, as they only differ in their tarval. */
3641 h = HASH_PTR(node->attr.con.tv);
3642 h = 9*h + HASH_PTR(get_irn_mode(node));
3643 } else if (node->op == op_SymConst) {
3644 /* special value for const, as they only differ in their symbol. */
3645 h = HASH_PTR(node->attr.i.sym.type_p);
3646 h = 9*h + HASH_PTR(get_irn_mode(node));
3649 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
3650 h = irn_arity = get_irn_intra_arity(node);
3652 /* consider all in nodes... except the block if not a control flow. */
3653 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
3654 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
3658 h = 9*h + HASH_PTR(get_irn_mode(node));
3660 h = 9*h + HASH_PTR(get_irn_op(node));
3666 pset *new_identities(void) {
3667 return new_pset(identities_cmp, N_IR_NODES);
3670 void del_identities(pset *value_table) {
3671 del_pset(value_table);
3675 * Return the canonical node computing the same value as n.
3676 * Looks up the node in a hash table.
3678 * For Const nodes this is performed in the constructor, too. Const
3679 * nodes are extremely time critical because of their frequent use in
3680 * constant string arrays.
3682 static INLINE ir_node *identify(pset *value_table, ir_node *n)
3686 if (!value_table) return n;
3688 if (get_opt_reassociation()) {
3689 if (is_op_commutative(get_irn_op(n))) {
3690 ir_node *l = get_binop_left(n);
3691 ir_node *r = get_binop_right(n);
3693 /* for commutative operators perform a OP b == b OP a */
3694 if (get_irn_idx(l) > get_irn_idx(r)) {
3695 set_binop_left(n, r);
3696 set_binop_right(n, l);
3701 o = pset_find(value_table, n, ir_node_hash(n));
3710 * During construction we set the op_pin_state_pinned flag in the graph right when the
3711 * optimization is performed. The flag turning on procedure global cse could
3712 * be changed between two allocations. This way we are safe.
3714 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
3717 n = identify(value_table, n);
3718 if (get_irn_n(old, -1) != get_irn_n(n, -1))
3719 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3724 * Return the canonical node computing the same value as n.
3725 * Looks up the node in a hash table, enters it in the table
3726 * if it isn't there yet.
3728 ir_node *identify_remember(pset *value_table, ir_node *n)
3732 if (!value_table) return n;
3734 if (get_opt_reassociation()) {
3735 if (is_op_commutative(get_irn_op(n))) {
3736 ir_node *l = get_binop_left(n);
3737 ir_node *r = get_binop_right(n);
3739 /* for commutative operators perform a OP b == b OP a */
3741 set_binop_left(n, r);
3742 set_binop_right(n, l);
3747 /* lookup or insert in hash table with given hash key. */
3748 o = pset_insert (value_table, n, ir_node_hash (n));
3757 /* Add a node to the identities value table. */
3758 void add_identities(pset *value_table, ir_node *node) {
3759 if (get_opt_cse() && is_no_Block(node))
3760 identify_remember(value_table, node);
3763 /* Visit each node in the value table of a graph. */
3764 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
3766 ir_graph *rem = current_ir_graph;
3768 current_ir_graph = irg;
3769 foreach_pset(irg->value_table, node)
3771 current_ir_graph = rem;
3775 * Garbage in, garbage out. If a node has a dead input, i.e., the
3776 * Bad node is input to the node, return the Bad node.
3778 static INLINE ir_node *gigo(ir_node *node)
3781 ir_op *op = get_irn_op(node);
3783 /* remove garbage blocks by looking at control flow that leaves the block
3784 and replacing the control flow by Bad. */
3785 if (get_irn_mode(node) == mode_X) {
3786 ir_node *block = get_nodes_block(skip_Proj(node));
3788 /* Don't optimize nodes in immature blocks. */
3789 if (!get_Block_matured(block)) return node;
3790 /* Don't optimize End, may have Bads. */
3791 if (op == op_End) return node;
3793 if (is_Block(block)) {
3794 irn_arity = get_irn_arity(block);
3795 for (i = 0; i < irn_arity; i++) {
3796 if (!is_Bad(get_irn_n(block, i)))
3799 if (i == irn_arity) return new_Bad();
3803 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
3804 blocks predecessors is dead. */
3805 if (op != op_Block && op != op_Phi && op != op_Tuple) {
3806 irn_arity = get_irn_arity(node);
3809 * Beware: we can only read the block of a non-floating node.
3811 if (is_irn_pinned_in_irg(node) &&
3812 is_Block_dead(get_nodes_block(node)))
3815 for (i = 0; i < irn_arity; i++) {
3816 ir_node *pred = get_irn_n(node, i);
3821 /* Propagating Unknowns here seems to be a bad idea, because
3822 sometimes we need a node as a input and did not want that
3824 However, it might be useful to move this into a later phase
3825 (if you think that optimizing such code is useful). */
3826 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
3827 return new_Unknown(get_irn_mode(node));
3832 /* With this code we violate the agreement that local_optimize
3833 only leaves Bads in Block, Phi and Tuple nodes. */
3834 /* If Block has only Bads as predecessors it's garbage. */
3835 /* If Phi has only Bads as predecessors it's garbage. */
3836 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
3837 irn_arity = get_irn_arity(node);
3838 for (i = 0; i < irn_arity; i++) {
3839 if (!is_Bad(get_irn_n(node, i))) break;
3841 if (i == irn_arity) node = new_Bad();
3848 * These optimizations deallocate nodes from the obstack.
3849 * It can only be called if it is guaranteed that no other nodes
3850 * reference this one, i.e., right after construction of a node.
3852 * current_ir_graph must be set to the graph of the node!
3854 ir_node *optimize_node(ir_node *n)
3858 opcode iro = get_irn_opcode(n);
3860 /* Always optimize Phi nodes: part of the construction. */
3861 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
3863 /* constant expression evaluation / constant folding */
3864 if (get_opt_constant_folding()) {
3865 /* neither constants nor Tuple values can be evaluated */
3866 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
3867 /* try to evaluate */
3868 tv = computed_value(n);
3869 if (tv != tarval_bad) {
3871 ir_type *old_tp = get_irn_type(n);
3872 int i, arity = get_irn_arity(n);
3876 * Try to recover the type of the new expression.
3878 for (i = 0; i < arity && !old_tp; ++i)
3879 old_tp = get_irn_type(get_irn_n(n, i));
3882 * we MUST copy the node here temporary, because it's still needed
3883 * for DBG_OPT_CSTEVAL
3885 node_size = offsetof(ir_node, attr) + n->op->attr_size;
3886 oldn = alloca(node_size);
3888 memcpy(oldn, n, node_size);
3889 CLONE_ARR_A(ir_node *, oldn->in, n->in);
3891 /* ARG, copy the in array, we need it for statistics */
3892 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
3894 /* note the inplace edges module */
3895 edges_node_deleted(n, current_ir_graph);
3897 /* evaluation was successful -- replace the node. */
3898 irg_kill_node(current_ir_graph, n);
3899 nw = new_Const(get_tarval_mode (tv), tv);
3901 if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv))
3902 set_Const_type(nw, old_tp);
3903 DBG_OPT_CSTEVAL(oldn, nw);
3909 /* remove unnecessary nodes */
3910 if (get_opt_constant_folding() ||
3911 (iro == iro_Phi) || /* always optimize these nodes. */
3913 (iro == iro_Proj) ||
3914 (iro == iro_Block) ) /* Flags tested local. */
3915 n = equivalent_node (n);
3917 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
3919 /* Common Subexpression Elimination.
3921 * Checks whether n is already available.
3922 * The block input is used to distinguish different subexpressions. Right
3923 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
3924 * subexpressions within a block.
3927 n = identify_cons (current_ir_graph->value_table, n);
3930 edges_node_deleted(oldn, current_ir_graph);
3932 /* We found an existing, better node, so we can deallocate the old node. */
3933 irg_kill_node(current_ir_graph, oldn);
3937 /* Some more constant expression evaluation that does not allow to
3939 iro = get_irn_opcode(n);
3940 if (get_opt_constant_folding() ||
3941 (iro == iro_Cond) ||
3942 (iro == iro_Proj) ||
3943 (iro == iro_Sel)) /* Flags tested local. */
3944 n = transform_node (n);
3946 /* Remove nodes with dead (Bad) input.
3947 Run always for transformation induced Bads. */
3950 /* Now we have a legal, useful node. Enter it in hash table for CSE */
3951 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
3952 n = identify_remember (current_ir_graph->value_table, n);
3960 * These optimizations never deallocate nodes (in place). This can cause dead
3961 * nodes lying on the obstack. Remove these by a dead node elimination,
3962 * i.e., a copying garbage collection.
3964 ir_node *optimize_in_place_2(ir_node *n)
3968 opcode iro = get_irn_opcode(n);
3970 if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n;
3972 /* constant expression evaluation / constant folding */
3973 if (get_opt_constant_folding()) {
3974 /* neither constants nor Tuple values can be evaluated */
3975 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
3976 /* try to evaluate */
3977 tv = computed_value(n);
3978 if (tv != tarval_bad) {
3979 /* evaluation was successful -- replace the node. */
3980 ir_type *old_tp = get_irn_type(n);
3981 int i, arity = get_irn_arity(n);
3984 * Try to recover the type of the new expression.
3986 for (i = 0; i < arity && !old_tp; ++i)
3987 old_tp = get_irn_type(get_irn_n(n, i));
3989 n = new_Const(get_tarval_mode(tv), tv);
3991 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
3992 set_Const_type(n, old_tp);
3994 DBG_OPT_CSTEVAL(oldn, n);
4000 /* remove unnecessary nodes */
4001 if (get_opt_constant_folding() ||
4002 (iro == iro_Phi) || /* always optimize these nodes. */
4003 (iro == iro_Id) || /* ... */
4004 (iro == iro_Proj) || /* ... */
4005 (iro == iro_Block) ) /* Flags tested local. */
4006 n = equivalent_node(n);
4008 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
4010 /** common subexpression elimination **/
4011 /* Checks whether n is already available. */
4012 /* The block input is used to distinguish different subexpressions. Right
4013 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
4014 subexpressions within a block. */
4015 if (get_opt_cse()) {
4016 n = identify(current_ir_graph->value_table, n);
4019 /* Some more constant expression evaluation. */
4020 iro = get_irn_opcode(n);
4021 if (get_opt_constant_folding() ||
4022 (iro == iro_Cond) ||
4023 (iro == iro_Proj) ||
4024 (iro == iro_Sel)) /* Flags tested local. */
4025 n = transform_node(n);
4027 /* Remove nodes with dead (Bad) input.
4028 Run always for transformation induced Bads. */
4031 /* Now we can verify the node, as it has no dead inputs any more. */
4034 /* Now we have a legal, useful node. Enter it in hash table for cse.
4035 Blocks should be unique anyways. (Except the successor of start:
4036 is cse with the start block!) */
4037 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
4038 n = identify_remember(current_ir_graph->value_table, n);
4044 * Wrapper for external use, set proper status bits after optimization.
4046 ir_node *optimize_in_place(ir_node *n)
4048 /* Handle graph state */
4049 assert(get_irg_phase_state(current_ir_graph) != phase_building);
4051 if (get_opt_global_cse())
4052 set_irg_pinned(current_ir_graph, op_pin_state_floats);
4053 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
4054 set_irg_outs_inconsistent(current_ir_graph);
4056 /* FIXME: Maybe we could also test whether optimizing the node can
4057 change the control graph. */
4058 set_irg_doms_inconsistent(current_ir_graph);
4059 return optimize_in_place_2 (n);
4063 * Sets the default operation for an ir_ops.
4065 ir_op_ops *firm_set_default_operations(opcode code, ir_op_ops *ops)
4067 ops = firm_set_default_computed_value(code, ops);
4068 ops = firm_set_default_equivalent_node(code, ops);
4069 ops = firm_set_default_transform_node(code, ops);
4070 ops = firm_set_default_node_cmp_attr(code, ops);
4071 ops = firm_set_default_get_type(code, ops);
4072 ops = firm_set_default_get_type_attr(code, ops);
4073 ops = firm_set_default_get_entity_attr(code, ops);