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;
900 ir_mode *mode = get_irn_mode(n);
902 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
903 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
906 n = equivalent_node_neutral_zero(n);
910 left = get_Add_left(n);
911 right = get_Add_right(n);
913 if (get_irn_op(left) == op_Sub) {
914 if (get_Sub_right(left) == right) {
917 n = get_Sub_left(left);
918 if (mode == get_irn_mode(n)) {
919 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
924 if (get_irn_op(right) == op_Sub) {
925 if (get_Sub_right(right) == left) {
928 n = get_Sub_left(right);
929 if (mode == get_irn_mode(n)) {
930 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
939 * optimize operations that are not commutative but have neutral 0 on left,
942 static ir_node *equivalent_node_left_zero(ir_node *n)
946 ir_node *a = get_binop_left(n);
947 ir_node *b = get_binop_right(n);
949 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
952 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
958 #define equivalent_node_Shl equivalent_node_left_zero
959 #define equivalent_node_Shr equivalent_node_left_zero
960 #define equivalent_node_Shrs equivalent_node_left_zero
961 #define equivalent_node_Rot equivalent_node_left_zero
964 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
966 * The second one looks strange, but this construct
967 * is used heavily in the LCC sources :-).
969 * Beware: The Mode of a Sub may be different than the mode of its
970 * predecessors, so we could not return a predecessors in all cases.
972 static ir_node *equivalent_node_Sub(ir_node *n)
976 ir_mode *mode = get_irn_mode(n);
978 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
979 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
983 b = get_Sub_right(n);
985 /* Beware: modes might be different */
986 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
987 if (mode == get_irn_mode(a)) {
990 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
993 else if (get_irn_op(a) == op_Add) {
994 if (mode_wrap_around(mode)) {
995 ir_node *left = get_Add_left(a);
996 ir_node *right = get_Add_right(a);
999 if (mode == get_irn_mode(right)) {
1001 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
1004 else if (right == b) {
1005 if (mode == get_irn_mode(left)) {
1007 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
1017 * Optimize an "idempotent unary op", ie op(op(n)) = n.
1020 * -(-a) == a, but might overflow two times.
1021 * We handle it anyway here but the better way would be a
1022 * flag. This would be needed for Pascal for instance.
1024 static ir_node *equivalent_node_idempotent_unop(ir_node *n)
1027 ir_node *pred = get_unop_op(n);
1029 /* optimize symmetric unop */
1030 if (get_irn_op(pred) == get_irn_op(n)) {
1031 n = get_unop_op(pred);
1032 DBG_OPT_ALGSIM2(oldn, pred, n);
1037 /** Not(Not(x)) == x */
1038 #define equivalent_node_Not equivalent_node_idempotent_unop
1040 /** --x == x ??? Is this possible or can --x raise an
1041 out of bounds exception if min =! max? */
1042 #define equivalent_node_Minus equivalent_node_idempotent_unop
1045 * Optimize a * 1 = 1 * a = a.
1047 static ir_node *equivalent_node_Mul(ir_node *n)
1050 ir_node *a = get_Mul_left(n);
1051 ir_node *b = get_Mul_right(n);
1053 /* Mul is commutative and has again an other neutral element. */
1054 if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) {
1056 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1057 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) {
1059 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1065 * Optimize a / 1 = a.
1067 static ir_node *equivalent_node_Div(ir_node *n)
1069 ir_node *a = get_Div_left(n);
1070 ir_node *b = get_Div_right(n);
1072 /* Div is not commutative. */
1073 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1074 /* Turn Div into a tuple (mem, bad, a) */
1075 ir_node *mem = get_Div_mem(n);
1076 turn_into_tuple(n, pn_Div_max);
1077 set_Tuple_pred(n, pn_Div_M, mem);
1078 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
1079 set_Tuple_pred(n, pn_Div_res, a);
1085 * Optimize a / 1.0 = a.
1087 static ir_node *equivalent_node_Quot(ir_node *n) {
1088 ir_node *a = get_Div_left(n);
1089 ir_node *b = get_Div_right(n);
1091 /* Div is not commutative. */
1092 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* Quot(x, 1) == x */
1093 /* Turn Quot into a tuple (mem, bad, a) */
1094 ir_node *mem = get_Quot_mem(n);
1095 turn_into_tuple(n, pn_Quot_max);
1096 set_Tuple_pred(n, pn_Quot_M, mem);
1097 set_Tuple_pred(n, pn_Quot_X_except, new_Bad()); /* no exception */
1098 set_Tuple_pred(n, pn_Quot_res, a);
1104 * Optimize a / 1 = a.
1106 static ir_node *equivalent_node_DivMod(ir_node *n)
1108 ir_node *a = get_DivMod_left(n);
1109 ir_node *b = get_DivMod_right(n);
1111 /* Div is not commutative. */
1112 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1113 /* Turn DivMod into a tuple (mem, bad, a, 0) */
1114 ir_node *mem = get_Div_mem(n);
1115 ir_mode *mode = get_irn_mode(b);
1117 turn_into_tuple(n, pn_DivMod_max);
1118 set_Tuple_pred(n, pn_DivMod_M, mem);
1119 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1120 set_Tuple_pred(n, pn_DivMod_res_div, a);
1121 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1127 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1129 static ir_node *equivalent_node_Or(ir_node *n)
1133 ir_node *a = get_Or_left(n);
1134 ir_node *b = get_Or_right(n);
1137 n = a; /* Or has it's own neutral element */
1138 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1139 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) {
1141 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1142 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
1144 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1151 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a.
1153 static ir_node *equivalent_node_And(ir_node *n)
1157 ir_node *a = get_And_left(n);
1158 ir_node *b = get_And_right(n);
1161 n = a; /* And has it's own neutral element */
1162 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1163 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) {
1165 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1166 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) {
1168 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1174 * Try to remove useless Conv's:
1176 static ir_node *equivalent_node_Conv(ir_node *n)
1179 ir_node *a = get_Conv_op(n);
1182 ir_mode *n_mode = get_irn_mode(n);
1183 ir_mode *a_mode = get_irn_mode(a);
1185 if (n_mode == a_mode) { /* No Conv necessary */
1186 /* leave strict floating point Conv's */
1187 if (get_Conv_strict(n))
1190 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1191 } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */
1195 n_mode = get_irn_mode(n);
1196 b_mode = get_irn_mode(b);
1198 if (n_mode == b_mode) {
1199 if (n_mode == mode_b) {
1200 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1201 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1203 else if (mode_is_int(n_mode) || mode_is_character(n_mode)) {
1204 if (smaller_mode(b_mode, a_mode)){
1205 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1206 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1215 * A Cast may be removed if the type of the previous node
1216 * is already the type of the Cast.
1218 static ir_node *equivalent_node_Cast(ir_node *n) {
1220 ir_node *pred = get_Cast_op(n);
1222 if (get_irn_type(pred) == get_Cast_type(n)) {
1224 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1230 Several optimizations:
1231 - no Phi in start block.
1232 - remove Id operators that are inputs to Phi
1233 - fold Phi-nodes, iff they have only one predecessor except
1236 static ir_node *equivalent_node_Phi(ir_node *n)
1241 ir_node *block = NULL; /* to shutup gcc */
1242 ir_node *first_val = NULL; /* to shutup gcc */
1244 if (!get_opt_normalize()) return n;
1246 n_preds = get_Phi_n_preds(n);
1248 block = get_nodes_block(n);
1249 /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
1250 assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
1251 if ((is_Block_dead(block)) || /* Control dead */
1252 (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */
1253 return new_Bad(); /* in the Start Block. */
1255 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1257 /* If the Block has a Bad pred, we also have one. */
1258 for (i = 0; i < n_preds; ++i)
1259 if (is_Bad(get_Block_cfgpred(block, i)))
1260 set_Phi_pred(n, i, new_Bad());
1262 /* Find first non-self-referencing input */
1263 for (i = 0; i < n_preds; ++i) {
1264 first_val = get_Phi_pred(n, i);
1265 if ( (first_val != n) /* not self pointer */
1267 && (! is_Bad(first_val))
1269 ) { /* value not dead */
1270 break; /* then found first value. */
1275 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1279 /* search for rest of inputs, determine if any of these
1280 are non-self-referencing */
1281 while (++i < n_preds) {
1282 ir_node *scnd_val = get_Phi_pred(n, i);
1283 if ( (scnd_val != n)
1284 && (scnd_val != first_val)
1286 && (! is_Bad(scnd_val))
1294 /* Fold, if no multiple distinct non-self-referencing inputs */
1296 DBG_OPT_PHI(oldn, n);
1302 Several optimizations:
1303 - no Sync in start block.
1304 - fold Sync-nodes, iff they have only one predecessor except
1307 static ir_node *equivalent_node_Sync(ir_node *n)
1312 ir_node *first_val = NULL; /* to shutup gcc */
1314 if (!get_opt_normalize()) return n;
1316 n_preds = get_Sync_n_preds(n);
1318 /* Find first non-self-referencing input */
1319 for (i = 0; i < n_preds; ++i) {
1320 first_val = get_Sync_pred(n, i);
1321 if ((first_val != n) /* not self pointer */ &&
1322 (! is_Bad(first_val))
1323 ) { /* value not dead */
1324 break; /* then found first value. */
1329 /* A totally Bad or self-referencing Sync (we didn't break the above loop) */
1332 /* search the rest of inputs, determine if any of these
1333 are non-self-referencing */
1334 while (++i < n_preds) {
1335 ir_node *scnd_val = get_Sync_pred(n, i);
1336 if ((scnd_val != n) &&
1337 (scnd_val != first_val) &&
1338 (! is_Bad(scnd_val))
1344 /* Fold, if no multiple distinct non-self-referencing inputs */
1346 DBG_OPT_SYNC(oldn, n);
1352 * optimize Proj(Tuple) and gigo() for ProjX in Bad block,
1353 * ProjX(Load) and ProjX(Store)
1355 static ir_node *equivalent_node_Proj(ir_node *n)
1359 ir_node *a = get_Proj_pred(n);
1361 if ( get_irn_op(a) == op_Tuple) {
1362 /* Remove the Tuple/Proj combination. */
1363 if ( get_Proj_proj(n) <= get_Tuple_n_preds(a) ) {
1364 n = get_Tuple_pred(a, get_Proj_proj(n));
1365 DBG_OPT_TUPLE(oldn, a, n);
1367 assert(0); /* This should not happen! */
1371 else if (get_irn_mode(n) == mode_X) {
1372 if (is_Block_dead(get_nodes_block(skip_Proj(n)))) {
1373 /* Remove dead control flow -- early gigo(). */
1376 else if (get_opt_ldst_only_null_ptr_exceptions()) {
1377 ir_op *op = get_irn_op(a);
1379 if (op == op_Load || op == op_Store) {
1380 /* get the load/store address */
1381 ir_node *addr = get_irn_n(a, 1);
1382 if (value_not_null(addr)) {
1383 /* this node may float if it did not depend on a Confirm */
1384 set_irn_pinned(a, op_pin_state_floats);
1398 static ir_node *equivalent_node_Id(ir_node *n)
1404 } while (get_irn_op(n) == op_Id);
1406 DBG_OPT_ID(oldn, n);
1413 static ir_node *equivalent_node_Mux(ir_node *n)
1415 ir_node *oldn = n, *sel = get_Mux_sel(n);
1416 tarval *ts = value_of(sel);
1418 /* Mux(true, f, t) == t */
1419 if (ts == tarval_b_true) {
1420 n = get_Mux_true(n);
1421 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1423 /* Mux(false, f, t) == f */
1424 else if (ts == tarval_b_false) {
1425 n = get_Mux_false(n);
1426 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1428 /* Mux(v, x, x) == x */
1429 else if (get_Mux_false(n) == get_Mux_true(n)) {
1430 n = get_Mux_true(n);
1431 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1433 else if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(get_irn_mode(n))) {
1434 ir_node *cmp = get_Proj_pred(sel);
1435 long proj_nr = get_Proj_proj(sel);
1436 ir_node *b = get_Mux_false(n);
1437 ir_node *a = get_Mux_true(n);
1440 * Note: normalization puts the constant on the right site,
1441 * so we check only one case.
1443 * Note further that these optimization work even for floating point
1444 * with NaN's because -NaN == NaN.
1445 * However, if +0 and -0 is handled differently, we cannot use the first one.
1447 if (get_irn_op(cmp) == op_Cmp && get_Cmp_left(cmp) == a) {
1448 if (classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
1449 /* Mux(a CMP 0, X, a) */
1450 if (get_irn_op(b) == op_Minus && get_Minus_op(b) == a) {
1451 /* Mux(a CMP 0, -a, a) */
1452 if (proj_nr == pn_Cmp_Eq) {
1453 /* Mux(a == 0, -a, a) ==> -a */
1455 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1457 else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1458 /* Mux(a != 0, -a, a) ==> a */
1460 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1463 else if (classify_Const(b) == CNST_NULL) {
1464 /* Mux(a CMP 0, 0, a) */
1465 if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1466 /* Mux(a != 0, 0, a) ==> a */
1468 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1470 else if (proj_nr == pn_Cmp_Eq) {
1471 /* Mux(a == 0, 0, a) ==> 0 */
1473 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1483 * Returns a equivalent node of a Psi: if a condition is true
1484 * and all previous conditions are false we know its value.
1485 * If all conditions are false its value is the default one.
1487 static ir_node *equivalent_node_Psi(ir_node *n) {
1489 return equivalent_node_Mux(n);
1494 * Optimize -a CMP -b into b CMP a.
1495 * This works only for for modes where unary Minus
1497 * Note that two-complement integers can Overflow
1498 * so it will NOT work.
1500 static ir_node *equivalent_node_Cmp(ir_node *n)
1502 ir_node *left = get_Cmp_left(n);
1503 ir_node *right = get_Cmp_right(n);
1505 if (get_irn_op(left) == op_Minus && get_irn_op(right) == op_Minus &&
1506 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1507 left = get_Minus_op(left);
1508 right = get_Minus_op(right);
1509 set_Cmp_left(n, right);
1510 set_Cmp_right(n, left);
1516 * Remove Confirm nodes if setting is on.
1517 * Replace Confirms(x, '=', Constlike) by Constlike.
1519 static ir_node *equivalent_node_Confirm(ir_node *n)
1521 ir_node *pred = get_Confirm_value(n);
1522 pn_Cmp pnc = get_Confirm_cmp(n);
1524 if (get_irn_op(pred) == op_Confirm && pnc == get_Confirm_cmp(pred)) {
1526 * rare case: two identical Confirms one after another,
1527 * replace the second one with the first.
1531 if (pnc == pn_Cmp_Eq) {
1532 ir_node *bound = get_Confirm_bound(n);
1535 * Optimize a rare case:
1536 * Confirm(x, '=', Constlike) ==> Constlike
1538 if (is_irn_constlike(bound)) {
1539 DBG_OPT_CONFIRM(n, bound);
1543 return get_opt_remove_confirm() ? get_Confirm_value(n) : n;
1547 * Optimize CopyB(mem, x, x) into a Nop
1549 static ir_node *equivalent_node_CopyB(ir_node *n)
1551 ir_node *a = get_CopyB_dst(n);
1552 ir_node *b = get_CopyB_src(n);
1555 /* Turn CopyB into a tuple (mem, bad, bad) */
1556 ir_node *mem = get_CopyB_mem(n);
1557 turn_into_tuple(n, pn_CopyB_max);
1558 set_Tuple_pred(n, pn_CopyB_M, mem);
1559 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1560 set_Tuple_pred(n, pn_CopyB_M_except, new_Bad());
1566 * Optimize Bounds(idx, idx, upper) into idx.
1568 static ir_node *equivalent_node_Bound(ir_node *n)
1570 ir_node *idx = get_Bound_index(n);
1571 ir_node *lower = get_Bound_lower(n);
1574 /* By definition lower < upper, so if idx == lower -->
1575 lower <= idx && idx < upper */
1577 /* Turn Bound into a tuple (mem, bad, idx) */
1581 ir_node *pred = skip_Proj(idx);
1583 if (get_irn_op(pred) == op_Bound) {
1585 * idx was Bounds_check previously, it is still valid if
1586 * lower <= pred_lower && pred_upper <= upper.
1588 ir_node *upper = get_Bound_upper(n);
1589 if (get_Bound_lower(pred) == lower &&
1590 get_Bound_upper(pred) == upper) {
1592 * One could expect that we simply return the previous
1593 * Bound here. However, this would be wrong, as we could
1594 * add an exception Proj to a new location than.
1595 * So, we must turn in into a tuple
1602 /* Turn Bound into a tuple (mem, bad, idx) */
1603 ir_node *mem = get_Bound_mem(n);
1604 turn_into_tuple(n, pn_Bound_max);
1605 set_Tuple_pred(n, pn_Bound_M, mem);
1606 set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */
1607 set_Tuple_pred(n, pn_Bound_res, idx);
1613 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1614 * perform no actual computation, as, e.g., the Id nodes. It does not create
1615 * new nodes. It is therefore safe to free n if the node returned is not n.
1616 * If a node returns a Tuple we can not just skip it. If the size of the
1617 * in array fits, we transform n into a tuple (e.g., Div).
1620 equivalent_node(ir_node *n)
1622 if (n->op->ops.equivalent_node)
1623 return n->op->ops.equivalent_node(n);
1628 * sets the default equivalent node operation for an ir_op_ops.
1630 * @param code the opcode for the default operation
1631 * @param ops the operations initialized
1636 static ir_op_ops *firm_set_default_equivalent_node(opcode code, ir_op_ops *ops)
1640 ops->equivalent_node = equivalent_node_##a; \
1683 * Do node specific optimizations of nodes predecessors.
1686 optimize_preds(ir_node *n) {
1687 ir_node *a = NULL, *b = NULL;
1689 /* get the operands we will work on for simple cases. */
1691 a = get_binop_left(n);
1692 b = get_binop_right(n);
1693 } else if (is_unop(n)) {
1697 switch (get_irn_opcode(n)) {
1700 /* We don't want Cast as input to Cmp. */
1701 if (get_irn_op(a) == op_Cast) {
1705 if (get_irn_op(b) == op_Cast) {
1707 set_Cmp_right(n, b);
1716 * Returns non-zero if a node is a Phi node
1717 * with all predecessors constant.
1719 static int is_const_Phi(ir_node *n) {
1724 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1725 if (! is_Const(get_irn_n(n, i)))
1731 * Apply an evaluator on a binop with a constant operators (and one Phi).
1733 * @param phi the Phi node
1734 * @param other the other operand
1735 * @param eval an evaluator function
1736 * @param left if non-zero, other is the left operand, else the right
1738 * @return a new Phi node if the conversion was successful, NULL else
1740 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), int left) {
1746 int i, n = get_irn_arity(phi);
1748 NEW_ARR_A(void *, res, n);
1750 for (i = 0; i < n; ++i) {
1751 pred = get_irn_n(phi, i);
1752 tv = get_Const_tarval(pred);
1753 tv = eval(other, tv);
1755 if (tv == tarval_bad) {
1756 /* folding failed, bad */
1763 for (i = 0; i < n; ++i) {
1764 pred = get_irn_n(phi, i);
1765 tv = get_Const_tarval(pred);
1766 tv = eval(tv, other);
1768 if (tv == tarval_bad) {
1769 /* folding failed, bad */
1775 mode = get_irn_mode(phi);
1776 irg = current_ir_graph;
1777 for (i = 0; i < n; ++i) {
1778 pred = get_irn_n(phi, i);
1779 res[i] = new_r_Const_type(irg, get_nodes_block(pred),
1780 mode, res[i], get_Const_type(pred));
1782 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1786 * Apply an evaluator on a unop with a constant operator (a Phi).
1788 * @param phi the Phi node
1789 * @param eval an evaluator function
1791 * @return a new Phi node if the conversion was successful, NULL else
1793 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
1799 int i, n = get_irn_arity(phi);
1801 NEW_ARR_A(void *, res, n);
1802 for (i = 0; i < n; ++i) {
1803 pred = get_irn_n(phi, i);
1804 tv = get_Const_tarval(pred);
1807 if (tv == tarval_bad) {
1808 /* folding failed, bad */
1813 mode = get_irn_mode(phi);
1814 irg = current_ir_graph;
1815 for (i = 0; i < n; ++i) {
1816 pred = get_irn_n(phi, i);
1817 res[i] = new_r_Const_type(irg, get_nodes_block(pred),
1818 mode, res[i], get_Const_type(pred));
1820 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1824 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1825 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1826 * If possible, remove the Conv's.
1828 static ir_node *transform_node_AddSub(ir_node *n)
1830 ir_mode *mode = get_irn_mode(n);
1832 if (mode_is_reference(mode)) {
1833 ir_node *left = get_binop_left(n);
1834 ir_node *right = get_binop_right(n);
1835 int ref_bits = get_mode_size_bits(mode);
1837 if (get_irn_op(left) == op_Conv) {
1838 ir_mode *mode = get_irn_mode(left);
1839 int bits = get_mode_size_bits(mode);
1841 if (ref_bits == bits &&
1842 mode_is_int(mode) &&
1843 get_mode_arithmetic(mode) == irma_twos_complement) {
1844 ir_node *pre = get_Conv_op(left);
1845 ir_mode *pre_mode = get_irn_mode(pre);
1847 if (mode_is_int(pre_mode) &&
1848 get_mode_size_bits(pre_mode) == bits &&
1849 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1850 /* ok, this conv just changes to sign, moreover the calculation
1851 * is done with same number of bits as our address mode, so
1852 * we can ignore the conv as address calculation can be viewed
1853 * as either signed or unsigned
1855 set_binop_left(n, pre);
1860 if (get_irn_op(right) == op_Conv) {
1861 ir_mode *mode = get_irn_mode(right);
1862 int bits = get_mode_size_bits(mode);
1864 if (ref_bits == bits &&
1865 mode_is_int(mode) &&
1866 get_mode_arithmetic(mode) == irma_twos_complement) {
1867 ir_node *pre = get_Conv_op(right);
1868 ir_mode *pre_mode = get_irn_mode(pre);
1870 if (mode_is_int(pre_mode) &&
1871 get_mode_size_bits(pre_mode) == bits &&
1872 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1873 /* ok, this conv just changes to sign, moreover the calculation
1874 * is done with same number of bits as our address mode, so
1875 * we can ignore the conv as address calculation can be viewed
1876 * as either signed or unsigned
1878 set_binop_right(n, pre);
1886 #define HANDLE_BINOP_PHI(op,a,b,c) \
1888 if (is_Const(b) && is_const_Phi(a)) { \
1889 /* check for Op(Phi, Const) */ \
1890 c = apply_binop_on_phi(a, get_Const_tarval(b), op, 0); \
1892 else if (is_Const(a) && is_const_Phi(b)) { \
1893 /* check for Op(Const, Phi) */ \
1894 c = apply_binop_on_phi(b, get_Const_tarval(a), op, 1); \
1897 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1901 #define HANDLE_UNOP_PHI(op,a,c) \
1903 if (is_const_Phi(a)) { \
1904 /* check for Op(Phi) */ \
1905 c = apply_unop_on_phi(a, op); \
1908 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1914 * Do the AddSub optimization, then Transform
1915 * Constant folding on Phi
1916 * Add(a,a) -> Mul(a, 2)
1917 * Add(Mul(a, x), a) -> Mul(a, x+1)
1918 * if the mode is integer or float.
1919 * Transform Add(a,-b) into Sub(a,b).
1920 * Reassociation might fold this further.
1922 static ir_node *transform_node_Add(ir_node *n)
1925 ir_node *a, *b, *c, *oldn = n;
1927 n = transform_node_AddSub(n);
1929 a = get_Add_left(n);
1930 b = get_Add_right(n);
1932 HANDLE_BINOP_PHI(tarval_add, a,b,c);
1934 mode = get_irn_mode(n);
1936 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1937 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1940 if (mode_is_num(mode)) {
1942 ir_node *block = get_irn_n(n, -1);
1945 get_irn_dbg_info(n),
1949 new_r_Const_long(current_ir_graph, block, mode, 2),
1951 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
1953 else if (get_irn_op(a) == op_Minus) {
1955 get_irn_dbg_info(n),
1961 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1963 else if (get_irn_op(b) == op_Minus) {
1965 get_irn_dbg_info(n),
1971 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1973 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1974 else if (!get_opt_reassociation() && get_irn_op(a) == op_Mul) {
1975 ir_node *ma = get_Mul_left(a);
1976 ir_node *mb = get_Mul_right(a);
1979 ir_node *blk = get_irn_n(n, -1);
1981 get_irn_dbg_info(n), current_ir_graph, blk,
1984 get_irn_dbg_info(n), current_ir_graph, blk,
1986 new_r_Const_long(current_ir_graph, blk, mode, 1),
1989 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1992 ir_node *blk = get_irn_n(n, -1);
1994 get_irn_dbg_info(n), current_ir_graph, blk,
1997 get_irn_dbg_info(n), current_ir_graph, blk,
1999 new_r_Const_long(current_ir_graph, blk, mode, 1),
2002 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
2005 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2006 else if (!get_opt_reassociation() && get_irn_op(b) == op_Mul) {
2007 ir_node *ma = get_Mul_left(b);
2008 ir_node *mb = get_Mul_right(b);
2011 ir_node *blk = get_irn_n(n, -1);
2013 get_irn_dbg_info(n), current_ir_graph, blk,
2016 get_irn_dbg_info(n), current_ir_graph, blk,
2018 new_r_Const_long(current_ir_graph, blk, mode, 1),
2021 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
2024 ir_node *blk = get_irn_n(n, -1);
2026 get_irn_dbg_info(n), current_ir_graph, blk,
2029 get_irn_dbg_info(n), current_ir_graph, blk,
2031 new_r_Const_long(current_ir_graph, blk, mode, 1),
2034 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
2042 * Do the AddSub optimization, then Transform
2043 * Constant folding on Phi
2044 * Sub(0,a) -> Minus(a)
2045 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2046 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2048 static ir_node *transform_node_Sub(ir_node *n)
2054 n = transform_node_AddSub(n);
2056 a = get_Sub_left(n);
2057 b = get_Sub_right(n);
2059 HANDLE_BINOP_PHI(tarval_sub, a,b,c);
2061 mode = get_irn_mode(n);
2063 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2064 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2067 if (mode_is_num(mode) && (classify_Const(a) == CNST_NULL)) {
2069 get_irn_dbg_info(n),
2074 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2076 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2077 else if (get_opt_reassociation() && get_irn_op(a) == op_Mul) {
2078 ir_node *ma = get_Mul_left(a);
2079 ir_node *mb = get_Mul_right(a);
2082 ir_node *blk = get_irn_n(n, -1);
2084 get_irn_dbg_info(n),
2085 current_ir_graph, blk,
2088 get_irn_dbg_info(n),
2089 current_ir_graph, blk,
2091 new_r_Const_long(current_ir_graph, blk, mode, 1),
2094 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2097 ir_node *blk = get_irn_n(n, -1);
2099 get_irn_dbg_info(n),
2100 current_ir_graph, blk,
2103 get_irn_dbg_info(n),
2104 current_ir_graph, blk,
2106 new_r_Const_long(current_ir_graph, blk, mode, 1),
2109 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2112 else if (get_irn_op(a) == op_Sub) {
2113 ir_node *x = get_Sub_left(a);
2114 ir_node *y = get_Sub_right(a);
2115 ir_node *blk = get_irn_n(n, -1);
2116 ir_mode *m_b = get_irn_mode(b);
2117 ir_mode *m_y = get_irn_mode(y);
2120 /* Determine the right mode for the Add. */
2123 else if (mode_is_reference(m_b))
2125 else if (mode_is_reference(m_y))
2129 * Both modes are different but none is reference,
2130 * happens for instance in SubP(SubP(P, Iu), Is).
2131 * We have two possibilities here: Cast or ignore.
2132 * Currently we ignore this case.
2137 add = new_r_Add(current_ir_graph, blk, y, b, mode);
2140 set_Sub_right(n, add);
2141 DBG_OPT_ALGSIM0(n, n, FS_OPT_SUB_SUB_X_Y_Z);
2148 * Transform Mul(a,-1) into -a.
2149 * Do constant evaluation of Phi nodes.
2150 * Do architecture dependent optimizations on Mul nodes
2152 static ir_node *transform_node_Mul(ir_node *n) {
2153 ir_node *c, *oldn = n;
2154 ir_node *a = get_Mul_left(n);
2155 ir_node *b = get_Mul_right(n);
2158 HANDLE_BINOP_PHI(tarval_mul, a,b,c);
2160 mode = get_irn_mode(n);
2161 if (mode_is_signed(mode)) {
2164 if (value_of(a) == get_mode_minus_one(mode))
2166 else if (value_of(b) == get_mode_minus_one(mode))
2169 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2170 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2174 return arch_dep_replace_mul_with_shifts(n);
2178 * Transform a Div Node.
2180 static ir_node *transform_node_Div(ir_node *n)
2182 tarval *tv = value_of(n);
2185 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2187 if (tv != tarval_bad) {
2188 value = new_Const(get_tarval_mode(tv), tv);
2190 DBG_OPT_CSTEVAL(n, value);
2192 else /* Try architecture dependent optimization */
2193 value = arch_dep_replace_div_by_const(n);
2196 /* Turn Div into a tuple (mem, bad, value) */
2197 ir_node *mem = get_Div_mem(n);
2199 turn_into_tuple(n, pn_Div_max);
2200 set_Tuple_pred(n, pn_Div_M, mem);
2201 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2202 set_Tuple_pred(n, pn_Div_res, value);
2208 * Transform a Mod node.
2210 static ir_node *transform_node_Mod(ir_node *n)
2212 tarval *tv = value_of(n);
2215 /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */
2217 if (tv != tarval_bad) {
2218 value = new_Const(get_tarval_mode(tv), tv);
2220 DBG_OPT_CSTEVAL(n, value);
2222 else /* Try architecture dependent optimization */
2223 value = arch_dep_replace_mod_by_const(n);
2226 /* Turn Mod into a tuple (mem, bad, value) */
2227 ir_node *mem = get_Mod_mem(n);
2229 turn_into_tuple(n, pn_Mod_max);
2230 set_Tuple_pred(n, pn_Mod_M, mem);
2231 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2232 set_Tuple_pred(n, pn_Mod_res, value);
2238 * Transform a DivMod node.
2240 static ir_node *transform_node_DivMod(ir_node *n)
2244 ir_node *a = get_DivMod_left(n);
2245 ir_node *b = get_DivMod_right(n);
2246 ir_mode *mode = get_irn_mode(a);
2247 tarval *ta = value_of(a);
2248 tarval *tb = value_of(b);
2250 if (!(mode_is_int(mode) && mode_is_int(get_irn_mode(b))))
2253 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2255 if (tb != tarval_bad) {
2256 if (tb == get_mode_one(get_tarval_mode(tb))) {
2257 b = new_Const (mode, get_mode_null(mode));
2260 DBG_OPT_CSTEVAL(n, b);
2262 else if (ta != tarval_bad) {
2263 tarval *resa, *resb;
2264 resa = tarval_div (ta, tb);
2265 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2266 Jmp for X result!? */
2267 resb = tarval_mod (ta, tb);
2268 if (resb == tarval_bad) return n; /* Causes exception! */
2269 a = new_Const (mode, resa);
2270 b = new_Const (mode, resb);
2273 DBG_OPT_CSTEVAL(n, a);
2274 DBG_OPT_CSTEVAL(n, b);
2276 else { /* Try architecture dependent optimization */
2277 arch_dep_replace_divmod_by_const(&a, &b, n);
2278 evaluated = a != NULL;
2280 } else if (ta == get_mode_null(mode)) {
2281 /* 0 / non-Const = 0 */
2286 if (evaluated) { /* replace by tuple */
2287 ir_node *mem = get_DivMod_mem(n);
2288 turn_into_tuple(n, pn_DivMod_max);
2289 set_Tuple_pred(n, pn_DivMod_M, mem);
2290 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
2291 set_Tuple_pred(n, pn_DivMod_res_div, a);
2292 set_Tuple_pred(n, pn_DivMod_res_mod, b);
2299 * Optimize Abs(x) into x if x is Confirmed >= 0
2300 * Optimize Abs(x) into -x if x is Confirmed <= 0
2302 static ir_node *transform_node_Abs(ir_node *n)
2305 ir_node *a = get_Abs_op(n);
2306 value_classify sign = classify_value_sign(a);
2308 if (sign == VALUE_NEGATIVE) {
2309 ir_mode *mode = get_irn_mode(n);
2312 * We can replace the Abs by -x here.
2313 * We even could add a new Confirm here.
2315 * Note that -x would create a new node, so we could
2316 * not run it in the equivalent_node() context.
2318 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
2319 get_irn_n(n, -1), a, mode);
2321 DBG_OPT_CONFIRM(oldn, n);
2323 else if (sign == VALUE_POSITIVE) {
2324 /* n is positive, Abs is not needed */
2327 DBG_OPT_CONFIRM(oldn, n);
2334 * Transform a Cond node.
2336 static ir_node *transform_node_Cond(ir_node *n)
2338 /* Replace the Cond by a Jmp if it branches on a constant
2341 ir_node *a = get_Cond_selector(n);
2342 tarval *ta = value_of(a);
2344 /* we need block info which is not available in floating irgs */
2345 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
2348 if ((ta != tarval_bad) &&
2349 (get_irn_mode(a) == mode_b) &&
2350 (get_opt_unreachable_code())) {
2351 /* It's a boolean Cond, branching on a boolean constant.
2352 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2353 jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n));
2354 turn_into_tuple(n, pn_Cond_max);
2355 if (ta == tarval_b_true) {
2356 set_Tuple_pred(n, pn_Cond_false, new_Bad());
2357 set_Tuple_pred(n, pn_Cond_true, jmp);
2359 set_Tuple_pred(n, pn_Cond_false, jmp);
2360 set_Tuple_pred(n, pn_Cond_true, new_Bad());
2362 /* We might generate an endless loop, so keep it alive. */
2363 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
2371 static ir_node *transform_node_And(ir_node *n)
2373 ir_node *c, *oldn = n;
2374 ir_node *a = get_And_left(n);
2375 ir_node *b = get_And_right(n);
2377 HANDLE_BINOP_PHI(tarval_and, a,b,c);
2384 static ir_node *transform_node_Eor(ir_node *n)
2386 ir_node *c, *oldn = n;
2387 ir_node *a = get_Eor_left(n);
2388 ir_node *b = get_Eor_right(n);
2389 ir_mode *mode = get_irn_mode(n);
2391 HANDLE_BINOP_PHI(tarval_eor, a,b,c);
2395 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
2396 mode, get_mode_null(mode));
2397 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
2399 else if ((mode == mode_b)
2400 && (get_irn_op(a) == op_Proj)
2401 && (get_irn_mode(a) == mode_b)
2402 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)
2403 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2404 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
2405 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2406 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
2408 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
2410 else if ((mode == mode_b)
2411 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) {
2412 /* The Eor is a Not. Replace it by a Not. */
2413 /* ????!!!Extend to bitfield 1111111. */
2414 n = new_r_Not(current_ir_graph, get_irn_n(n, -1), a, mode_b);
2416 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2425 static ir_node *transform_node_Not(ir_node *n)
2427 ir_node *c, *oldn = n;
2428 ir_node *a = get_Not_op(n);
2430 HANDLE_UNOP_PHI(tarval_not,a,c);
2432 /* check for a boolean Not */
2433 if ( (get_irn_mode(n) == mode_b)
2434 && (get_irn_op(a) == op_Proj)
2435 && (get_irn_mode(a) == mode_b)
2436 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2437 /* We negate a Cmp. The Cmp has the negated result anyways! */
2438 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2439 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
2440 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
2446 * Transform a Minus.
2448 static ir_node *transform_node_Minus(ir_node *n)
2450 ir_node *c, *oldn = n;
2451 ir_node *a = get_Minus_op(n);
2453 HANDLE_UNOP_PHI(tarval_neg,a,c);
2458 * Transform a Cast_type(Const) into a new Const_type
2460 static ir_node *transform_node_Cast(ir_node *n) {
2462 ir_node *pred = get_Cast_op(n);
2463 ir_type *tp = get_irn_type(n);
2465 if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) {
2466 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
2467 get_Const_tarval(pred), tp);
2468 DBG_OPT_CSTEVAL(oldn, n);
2469 } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) {
2470 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred),
2471 get_SymConst_kind(pred), tp);
2472 DBG_OPT_CSTEVAL(oldn, n);
2479 * Transform a Proj(Div) with a non-zero value.
2480 * Removes the exceptions and routes the memory to the NoMem node.
2482 static ir_node *transform_node_Proj_Div(ir_node *proj)
2484 ir_node *n = get_Proj_pred(proj);
2485 ir_node *b = get_Div_right(n);
2488 if (value_not_zero(b)) {
2489 /* div(x, y) && y != 0 */
2490 proj_nr = get_Proj_proj(proj);
2492 /* this node may float if it did not depend on a Confirm */
2493 set_irn_pinned(n, op_pin_state_floats);
2495 if (proj_nr == pn_Div_X_except) {
2496 /* we found an exception handler, remove it */
2497 DBG_OPT_EXC_REM(proj);
2500 else if (proj_nr == pn_Div_M) {
2501 ir_node *res = get_Div_mem(n);
2502 /* the memory Proj can only be removed if we divide by a
2503 real constant, but the node never produce a new memory */
2504 if (value_of(b) != tarval_bad) {
2505 /* this is a Div by a const, we can remove the memory edge */
2506 set_Div_mem(n, get_irg_no_mem(current_ir_graph));
2515 * Transform a Proj(Mod) with a non-zero value.
2516 * Removes the exceptions and routes the memory to the NoMem node.
2518 static ir_node *transform_node_Proj_Mod(ir_node *proj)
2520 ir_node *n = get_Proj_pred(proj);
2521 ir_node *b = get_Mod_right(n);
2524 if (value_not_zero(b)) {
2525 /* mod(x, y) && y != 0 */
2526 proj_nr = get_Proj_proj(proj);
2528 /* this node may float if it did not depend on a Confirm */
2529 set_irn_pinned(n, op_pin_state_floats);
2531 if (proj_nr == pn_Mod_X_except) {
2532 /* we found an exception handler, remove it */
2533 DBG_OPT_EXC_REM(proj);
2535 } else if (proj_nr == pn_Mod_M) {
2536 ir_node *res = get_Mod_mem(n);
2537 /* the memory Proj can only be removed if we divide by a
2538 real constant, but the node never produce a new memory */
2539 if (value_of(b) != tarval_bad) {
2540 /* this is a Mod by a const, we can remove the memory edge */
2541 set_Mod_mem(n, get_irg_no_mem(current_ir_graph));
2545 else if (proj_nr == pn_Mod_res && get_Mod_left(n) == b) {
2546 /* a % a = 0 if a != 0 */
2547 ir_mode *mode = get_irn_mode(proj);
2548 ir_node *res = new_Const(mode, get_mode_null(mode));
2550 DBG_OPT_CSTEVAL(n, res);
2558 * Transform a Proj(DivMod) with a non-zero value.
2559 * Removes the exceptions and routes the memory to the NoMem node.
2561 static ir_node *transform_node_Proj_DivMod(ir_node *proj)
2563 ir_node *n = get_Proj_pred(proj);
2564 ir_node *b = get_DivMod_right(n);
2567 if (value_not_zero(b)) {
2568 /* DivMod(x, y) && y != 0 */
2569 proj_nr = get_Proj_proj(proj);
2571 /* this node may float if it did not depend on a Confirm */
2572 set_irn_pinned(n, op_pin_state_floats);
2574 if (proj_nr == pn_DivMod_X_except) {
2575 /* we found an exception handler, remove it */
2576 DBG_OPT_EXC_REM(proj);
2579 else if (proj_nr == pn_DivMod_M) {
2580 ir_node *res = get_DivMod_mem(n);
2581 /* the memory Proj can only be removed if we divide by a
2582 real constant, but the node never produce a new memory */
2583 if (value_of(b) != tarval_bad) {
2584 /* this is a DivMod by a const, we can remove the memory edge */
2585 set_DivMod_mem(n, get_irg_no_mem(current_ir_graph));
2589 else if (proj_nr == pn_DivMod_res_mod && get_DivMod_left(n) == b) {
2590 /* a % a = 0 if a != 0 */
2591 ir_mode *mode = get_irn_mode(proj);
2592 ir_node *res = new_Const(mode, get_mode_null(mode));
2594 DBG_OPT_CSTEVAL(n, res);
2602 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
2604 static ir_node *transform_node_Proj_Cond(ir_node *proj)
2606 if (get_opt_unreachable_code()) {
2607 ir_node *n = get_Proj_pred(proj);
2608 ir_node *b = get_Cond_selector(n);
2610 if (mode_is_int(get_irn_mode(b))) {
2611 tarval *tb = value_of(b);
2613 if (tb != tarval_bad) {
2614 /* we have a constant switch */
2615 long num = get_Proj_proj(proj);
2617 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
2618 if (get_tarval_long(tb) == num) {
2619 /* Do NOT create a jump here, or we will have 2 control flow ops
2620 * in a block. This case is optimized away in optimize_cf(). */
2624 /* this case will NEVER be taken, kill it */
2635 * Normalizes and optimizes Cmp nodes.
2637 static ir_node *transform_node_Proj_Cmp(ir_node *proj)
2639 if (get_opt_reassociation()) {
2640 ir_node *n = get_Proj_pred(proj);
2641 ir_node *left = get_Cmp_left(n);
2642 ir_node *right = get_Cmp_right(n);
2646 ir_mode *mode = NULL;
2647 long proj_nr = get_Proj_proj(proj);
2650 * First step: normalize the compare op
2651 * by placing the constant on the right site
2652 * or moving the lower address node to the left.
2653 * We ignore the case that both are constants
2654 * this case should be optimized away.
2656 if (get_irn_op(right) == op_Const)
2658 else if (get_irn_op(left) == op_Const) {
2663 proj_nr = get_inversed_pnc(proj_nr);
2666 else if (get_irn_idx(left) > get_irn_idx(right)) {
2672 proj_nr = get_inversed_pnc(proj_nr);
2677 * Second step: Try to reduce the magnitude
2678 * of a constant. This may help to generate better code
2679 * later and may help to normalize more compares.
2680 * Of course this is only possible for integer values.
2683 mode = get_irn_mode(c);
2684 tv = get_Const_tarval(c);
2686 if (tv != tarval_bad) {
2687 /* the following optimization is possible on modes without Overflow
2688 * on Unary Minus or on == and !=:
2689 * -a CMP c ==> a swap(CMP) -c
2691 * Beware: for two-complement Overflow may occur, so only == and != can
2692 * be optimized, see this:
2693 * -MININT < 0 =/=> MININT > 0 !!!
2695 if (get_opt_constant_folding() && get_irn_op(left) == op_Minus &&
2696 (!mode_overflow_on_unary_Minus(mode) ||
2697 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
2698 left = get_Minus_op(left);
2699 tv = tarval_sub(get_mode_null(mode), tv);
2701 proj_nr = get_inversed_pnc(proj_nr);
2705 /* for integer modes, we have more */
2706 if (mode_is_int(mode)) {
2707 /* Ne includes Unordered which is not possible on integers.
2708 * However, frontends often use this wrong, so fix it here */
2709 if (proj_nr & pn_Cmp_Uo) {
2710 proj_nr &= ~pn_Cmp_Uo;
2711 set_Proj_proj(proj, proj_nr);
2714 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
2715 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
2716 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
2717 tv = tarval_sub(tv, get_mode_one(mode));
2719 proj_nr ^= pn_Cmp_Eq;
2722 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
2723 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
2724 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
2725 tv = tarval_add(tv, get_mode_one(mode));
2727 proj_nr ^= pn_Cmp_Eq;
2731 /* the following reassociations work only for == and != */
2732 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
2734 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
2735 if (classify_tarval(tv) == TV_CLASSIFY_NULL && get_irn_op(left) == op_Sub) {
2736 right = get_Sub_right(left);
2737 left = get_Sub_left(left);
2739 tv = value_of(right);
2743 if (tv != tarval_bad) {
2744 ir_op *op = get_irn_op(left);
2746 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
2748 ir_node *c1 = get_Sub_right(left);
2749 tarval *tv2 = value_of(c1);
2751 if (tv2 != tarval_bad) {
2752 tv2 = tarval_add(tv, value_of(c1));
2754 if (tv2 != tarval_bad) {
2755 left = get_Sub_left(left);
2761 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
2762 else if (op == op_Add) {
2763 ir_node *a_l = get_Add_left(left);
2764 ir_node *a_r = get_Add_right(left);
2768 if (get_irn_op(a_l) == op_Const) {
2770 tv2 = value_of(a_l);
2774 tv2 = value_of(a_r);
2777 if (tv2 != tarval_bad) {
2778 tv2 = tarval_sub(tv, tv2);
2780 if (tv2 != tarval_bad) {
2787 /* -a == c ==> a == -c, -a != c ==> a != -c */
2788 else if (op == op_Minus) {
2789 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
2791 if (tv2 != tarval_bad) {
2792 left = get_Minus_op(left);
2799 /* the following reassociations work only for <= */
2800 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2801 if (tv != tarval_bad) {
2802 ir_op *op = get_irn_op(left);
2804 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
2812 * optimization for AND:
2814 * And(x, C) == C ==> And(x, C) != 0
2815 * And(x, C) != C ==> And(X, C) == 0
2817 * if C is a single Bit constant.
2819 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) &&
2820 (get_irn_op(left) == op_And)) {
2821 if (is_single_bit_tarval(tv)) {
2822 /* check for Constant's match. We have check hare the tarvals,
2823 because our const might be changed */
2824 ir_node *la = get_And_left(left);
2825 ir_node *ra = get_And_right(left);
2826 if ((is_Const(la) && get_Const_tarval(la) == tv) ||
2827 (is_Const(ra) && get_Const_tarval(ra) == tv)) {
2828 /* fine: do the transformation */
2829 tv = get_mode_null(get_tarval_mode(tv));
2830 proj_nr ^= pn_Cmp_Leg;
2835 } /* tarval != bad */
2839 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
2841 if (changed & 2) /* need a new Const */
2842 right = new_Const(mode, tv);
2844 /* create a new compare */
2845 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block,
2848 set_Proj_pred(proj, n);
2849 set_Proj_proj(proj, proj_nr);
2856 * Does all optimizations on nodes that must be done on it's Proj's
2857 * because of creating new nodes.
2859 static ir_node *transform_node_Proj(ir_node *proj)
2861 ir_node *n = get_Proj_pred(proj);
2863 switch (get_irn_opcode(n)) {
2865 return transform_node_Proj_Div(proj);
2868 return transform_node_Proj_Mod(proj);
2871 return transform_node_Proj_DivMod(proj);
2874 return transform_node_Proj_Cond(proj);
2877 return transform_node_Proj_Cmp(proj);
2880 /* should not happen, but if it does will be optimized away */
2881 return equivalent_node_Proj(proj);
2890 * Move Confirms down through Phi nodes.
2892 static ir_node *transform_node_Phi(ir_node *phi) {
2894 ir_mode *mode = get_irn_mode(phi);
2896 if (mode_is_reference(mode)) {
2897 n = get_irn_arity(phi);
2899 /* Beware of Phi0 */
2901 ir_node *pred = get_irn_n(phi, 0);
2902 ir_node *bound, *new_Phi, *block, **in;
2905 if (! is_Confirm(pred))
2908 bound = get_Confirm_bound(pred);
2909 pnc = get_Confirm_cmp(pred);
2911 NEW_ARR_A(ir_node *, in, n);
2912 in[0] = get_Confirm_value(pred);
2914 for (i = 1; i < n; ++i) {
2915 pred = get_irn_n(phi, i);
2917 if (! is_Confirm(pred) ||
2918 get_Confirm_bound(pred) != bound ||
2919 get_Confirm_cmp(pred) != pnc)
2921 in[i] = get_Confirm_value(pred);
2923 /* move the Confirm nodes "behind" the Phi */
2924 block = get_irn_n(phi, -1);
2925 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
2926 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
2933 * returns the operands of a commutative bin-op, if one operand is
2934 * a const, it is returned as the second one.
2936 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c)
2938 ir_node *op_a = get_binop_left(binop);
2939 ir_node *op_b = get_binop_right(binop);
2941 assert(is_op_commutative(get_irn_op(binop)));
2943 if (get_irn_op(op_a) == op_Const) {
2954 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
2955 * Such pattern may arise in bitfield stores.
2957 * value c4 value c4 & c2
2958 * AND c3 AND c1 | c3
2963 static ir_node *transform_node_Or_bf_store(ir_node *or)
2967 ir_node *and_l, *c3;
2968 ir_node *value, *c4;
2969 ir_node *new_and, *new_const, *block;
2970 ir_mode *mode = get_irn_mode(or);
2972 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
2974 get_comm_Binop_Ops(or, &and, &c1);
2975 if ((get_irn_op(c1) != op_Const) || (get_irn_op(and) != op_And))
2978 get_comm_Binop_Ops(and, &or_l, &c2);
2979 if ((get_irn_op(c2) != op_Const) || (get_irn_op(or_l) != op_Or))
2982 get_comm_Binop_Ops(or_l, &and_l, &c3);
2983 if ((get_irn_op(c3) != op_Const) || (get_irn_op(and_l) != op_And))
2986 get_comm_Binop_Ops(and_l, &value, &c4);
2987 if (get_irn_op(c4) != op_Const)
2990 /* ok, found the pattern, check for conditions */
2991 assert(mode == get_irn_mode(and));
2992 assert(mode == get_irn_mode(or_l));
2993 assert(mode == get_irn_mode(and_l));
2995 tv1 = get_Const_tarval(c1);
2996 tv2 = get_Const_tarval(c2);
2997 tv3 = get_Const_tarval(c3);
2998 tv4 = get_Const_tarval(c4);
3000 tv = tarval_or(tv4, tv2);
3001 if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) {
3002 /* have at least one 0 at the same bit position */
3006 n_tv4 = tarval_not(tv4);
3007 if (tv3 != tarval_and(tv3, n_tv4)) {
3008 /* bit in the or_mask is outside the and_mask */
3012 n_tv2 = tarval_not(tv2);
3013 if (tv1 != tarval_and(tv1, n_tv2)) {
3014 /* bit in the or_mask is outside the and_mask */
3018 /* ok, all conditions met */
3019 block = get_irn_n(or, -1);
3021 new_and = new_r_And(current_ir_graph, block,
3022 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
3024 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
3026 set_Or_left(or, new_and);
3027 set_Or_right(or, new_const);
3029 /* check for more */
3030 return transform_node_Or_bf_store(or);
3034 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
3036 static ir_node *transform_node_Or_Rot(ir_node *or)
3038 ir_mode *mode = get_irn_mode(or);
3039 ir_node *shl, *shr, *block;
3040 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
3043 if (! mode_is_int(mode))
3046 shl = get_binop_left(or);
3047 shr = get_binop_right(or);
3049 if (get_irn_op(shl) == op_Shr) {
3050 if (get_irn_op(shr) != op_Shl)
3057 else if (get_irn_op(shl) != op_Shl)
3059 else if (get_irn_op(shr) != op_Shr)
3062 x = get_Shl_left(shl);
3063 if (x != get_Shr_left(shr))
3066 c1 = get_Shl_right(shl);
3067 c2 = get_Shr_right(shr);
3068 if (get_irn_op(c1) == op_Const && get_irn_op(c2) == op_Const) {
3069 tv1 = get_Const_tarval(c1);
3070 if (! tarval_is_long(tv1))
3073 tv2 = get_Const_tarval(c2);
3074 if (! tarval_is_long(tv2))
3077 if (get_tarval_long(tv1) + get_tarval_long(tv2)
3078 != get_mode_size_bits(mode))
3081 /* yet, condition met */
3082 block = get_irn_n(or, -1);
3084 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
3086 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
3089 else if (get_irn_op(c1) == op_Sub) {
3093 if (get_Sub_right(sub) != v)
3096 c1 = get_Sub_left(sub);
3097 if (get_irn_op(c1) != op_Const)
3100 tv1 = get_Const_tarval(c1);
3101 if (! tarval_is_long(tv1))
3104 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
3107 /* yet, condition met */
3108 block = get_nodes_block(or);
3110 /* a Rot right is not supported, so use a rot left */
3111 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
3113 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
3116 else if (get_irn_op(c2) == op_Sub) {
3120 c1 = get_Sub_left(sub);
3121 if (get_irn_op(c1) != op_Const)
3124 tv1 = get_Const_tarval(c1);
3125 if (! tarval_is_long(tv1))
3128 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
3131 /* yet, condition met */
3132 block = get_irn_n(or, -1);
3135 n = new_r_Rot(current_ir_graph, block, x, v, mode);
3137 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
3147 static ir_node *transform_node_Or(ir_node *n)
3149 ir_node *c, *oldn = n;
3150 ir_node *a = get_Or_left(n);
3151 ir_node *b = get_Or_right(n);
3153 HANDLE_BINOP_PHI(tarval_or, a,b,c);
3155 n = transform_node_Or_bf_store(n);
3156 n = transform_node_Or_Rot(n);
3163 static ir_node *transform_node(ir_node *n);
3166 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl.
3168 * Should be moved to reassociation?
3170 static ir_node *transform_node_shift(ir_node *n)
3172 ir_node *left, *right;
3173 tarval *tv1, *tv2, *res;
3175 int modulo_shf, flag;
3177 left = get_binop_left(n);
3179 /* different operations */
3180 if (get_irn_op(left) != get_irn_op(n))
3183 right = get_binop_right(n);
3184 tv1 = value_of(right);
3185 if (tv1 == tarval_bad)
3188 tv2 = value_of(get_binop_right(left));
3189 if (tv2 == tarval_bad)
3192 res = tarval_add(tv1, tv2);
3194 /* beware: a simple replacement works only, if res < modulo shift */
3195 mode = get_irn_mode(n);
3199 modulo_shf = get_mode_modulo_shift(mode);
3200 if (modulo_shf > 0) {
3201 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
3203 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
3210 /* ok, we can replace it */
3211 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
3213 in[0] = get_binop_left(left);
3214 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
3216 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
3218 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
3220 return transform_node(irn);
3228 static ir_node *transform_node_Shr(ir_node *n)
3230 ir_node *c, *oldn = n;
3231 ir_node *a = get_Shr_left(n);
3232 ir_node *b = get_Shr_right(n);
3234 HANDLE_BINOP_PHI(tarval_shr, a, b, c);
3235 return transform_node_shift(n);
3241 static ir_node *transform_node_Shrs(ir_node *n)
3243 ir_node *c, *oldn = n;
3244 ir_node *a = get_Shrs_left(n);
3245 ir_node *b = get_Shrs_right(n);
3247 HANDLE_BINOP_PHI(tarval_shrs, a, b, c);
3248 return transform_node_shift(n);
3254 static ir_node *transform_node_Shl(ir_node *n)
3256 ir_node *c, *oldn = n;
3257 ir_node *a = get_Shl_left(n);
3258 ir_node *b = get_Shl_right(n);
3260 HANDLE_BINOP_PHI(tarval_shl, a, b, c);
3261 return transform_node_shift(n);
3265 * Remove dead blocks and nodes in dead blocks
3266 * in keep alive list. We do not generate a new End node.
3268 static ir_node *transform_node_End(ir_node *n) {
3269 int i, n_keepalives = get_End_n_keepalives(n);
3271 for (i = 0; i < n_keepalives; ++i) {
3272 ir_node *ka = get_End_keepalive(n, i);
3274 if (is_Block_dead(ka)) {
3275 set_End_keepalive(n, i, new_Bad());
3278 else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka)))
3279 set_End_keepalive(n, i, new_Bad());
3285 * Optimize a Mux into some simpler cases.
3287 static ir_node *transform_node_Mux(ir_node *n)
3289 ir_node *oldn = n, *sel = get_Mux_sel(n);
3290 ir_mode *mode = get_irn_mode(n);
3292 if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(mode)) {
3293 ir_node *cmp = get_Proj_pred(sel);
3294 long proj_nr = get_Proj_proj(sel);
3295 ir_node *f = get_Mux_false(n);
3296 ir_node *t = get_Mux_true(n);
3298 if (get_irn_op(cmp) == op_Cmp && classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
3299 ir_node *block = get_irn_n(n, -1);
3302 * Note: normalization puts the constant on the right site,
3303 * so we check only one case.
3305 * Note further that these optimization work even for floating point
3306 * with NaN's because -NaN == NaN.
3307 * However, if +0 and -0 is handled differently, we cannot use the first one.
3309 if (get_irn_op(f) == op_Minus &&
3310 get_Minus_op(f) == t &&
3311 get_Cmp_left(cmp) == t) {
3313 if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3314 /* Mux(a >=/> 0, -a, a) ==> Abs(a) */
3315 n = new_rd_Abs(get_irn_dbg_info(n),
3319 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3322 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3323 /* Mux(a <=/< 0, -a, a) ==> Minus(Abs(a)) */
3324 n = new_rd_Abs(get_irn_dbg_info(n),
3328 n = new_rd_Minus(get_irn_dbg_info(n),
3333 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3337 else if (get_irn_op(t) == op_Minus &&
3338 get_Minus_op(t) == f &&
3339 get_Cmp_left(cmp) == f) {
3341 if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3342 /* Mux(a <=/< 0, a, -a) ==> Abs(a) */
3343 n = new_rd_Abs(get_irn_dbg_info(n),
3347 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3350 else if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3351 /* Mux(a >=/> 0, a, -a) ==> Minus(Abs(a)) */
3352 n = new_rd_Abs(get_irn_dbg_info(n),
3356 n = new_rd_Minus(get_irn_dbg_info(n),
3361 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3366 if (mode_is_int(mode) && mode_is_signed(mode) &&
3367 get_mode_arithmetic(mode) == irma_twos_complement) {
3368 ir_node *x = get_Cmp_left(cmp);
3370 /* the following optimization works only with signed integer two-complement mode */
3372 if (mode == get_irn_mode(x)) {
3374 * FIXME: this restriction is two rigid, as it would still
3375 * work if mode(x) = Hs and mode == Is, but at least it removes
3378 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Le) &&
3379 classify_Const(t) == CNST_ALL_ONE &&
3380 classify_Const(f) == CNST_NULL) {
3382 * Mux(x:T </<= 0, 0, -1) -> Shrs(x, sizeof_bits(T) - 1)
3386 n = new_rd_Shrs(get_irn_dbg_info(n),
3387 current_ir_graph, block, x,
3388 new_r_Const_long(current_ir_graph, block, mode_Iu,
3389 get_mode_size_bits(mode) - 1),
3391 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3394 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) &&
3395 classify_Const(t) == CNST_ONE &&
3396 classify_Const(f) == CNST_NULL) {
3398 * Mux(x:T >/>= 0, 0, 1) -> Shr(-x, sizeof_bits(T) - 1)
3402 n = new_rd_Shr(get_irn_dbg_info(n),
3403 current_ir_graph, block,
3404 new_r_Minus(current_ir_graph, block, x, mode),
3405 new_r_Const_long(current_ir_graph, block, mode_Iu,
3406 get_mode_size_bits(mode) - 1),
3408 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3415 return arch_transform_node_Mux(n);
3419 * Optimize a Psi into some simpler cases.
3421 static ir_node *transform_node_Psi(ir_node *n) {
3423 return transform_node_Mux(n);
3429 * Tries several [inplace] [optimizing] transformations and returns an
3430 * equivalent node. The difference to equivalent_node() is that these
3431 * transformations _do_ generate new nodes, and thus the old node must
3432 * not be freed even if the equivalent node isn't the old one.
3434 static ir_node *transform_node(ir_node *n)
3436 if (n->op->ops.transform_node)
3437 n = n->op->ops.transform_node(n);
3442 * sSets the default transform node operation for an ir_op_ops.
3444 * @param code the opcode for the default operation
3445 * @param ops the operations initialized
3450 static ir_op_ops *firm_set_default_transform_node(opcode code, ir_op_ops *ops)
3454 ops->transform_node = transform_node_##a; \
3490 /* **************** Common Subexpression Elimination **************** */
3492 /** The size of the hash table used, should estimate the number of nodes
3494 #define N_IR_NODES 512
3496 /** Compares the attributes of two Const nodes. */
3497 static int node_cmp_attr_Const(ir_node *a, ir_node *b)
3499 return (get_Const_tarval(a) != get_Const_tarval(b))
3500 || (get_Const_type(a) != get_Const_type(b));
3503 /** Compares the attributes of two Proj nodes. */
3504 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
3505 return get_irn_proj_attr (a) != get_irn_proj_attr (b);
3508 /** Compares the attributes of two Filter nodes. */
3509 static int node_cmp_attr_Filter(ir_node *a, ir_node *b)
3511 return get_Filter_proj(a) != get_Filter_proj(b);
3514 /** Compares the attributes of two Alloc nodes. */
3515 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b)
3517 return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where)
3518 || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type);
3521 /** Compares the attributes of two Free nodes. */
3522 static int node_cmp_attr_Free(ir_node *a, ir_node *b)
3524 return (get_irn_free_attr(a).where != get_irn_free_attr(b).where)
3525 || (get_irn_free_attr(a).type != get_irn_free_attr(b).type);
3528 /** Compares the attributes of two SymConst nodes. */
3529 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b)
3531 return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num)
3532 || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p)
3533 || (get_irn_symconst_attr(a).tp != get_irn_symconst_attr(b).tp);
3536 /** Compares the attributes of two Call nodes. */
3537 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
3538 return (get_irn_call_attr(a) != get_irn_call_attr(b));
3541 /** Compares the attributes of two Sel nodes. */
3542 static int node_cmp_attr_Sel(ir_node *a, ir_node *b)
3544 return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind)
3545 || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name)
3546 || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner)
3547 || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name)
3548 || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type);
3551 /** Compares the attributes of two Phi nodes. */
3552 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
3553 return get_irn_phi_attr (a) != get_irn_phi_attr (b);
3556 /** Compares the attributes of two Conv nodes. */
3557 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
3558 return get_Conv_strict(a) != get_Conv_strict(b);
3561 /** Compares the attributes of two Cast nodes. */
3562 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
3563 return get_Cast_type(a) != get_Cast_type(b);
3566 /** Compares the attributes of two Load nodes. */
3567 static int node_cmp_attr_Load(ir_node *a, ir_node *b)
3569 if (get_Load_volatility(a) == volatility_is_volatile ||
3570 get_Load_volatility(b) == volatility_is_volatile)
3571 /* NEVER do CSE on volatile Loads */
3574 return get_Load_mode(a) != get_Load_mode(b);
3577 /** Compares the attributes of two Store nodes. */
3578 static int node_cmp_attr_Store(ir_node *a, ir_node *b)
3580 /* NEVER do CSE on volatile Stores */
3581 return (get_Store_volatility(a) == volatility_is_volatile ||
3582 get_Store_volatility(b) == volatility_is_volatile);
3585 /** Compares the attributes of two Confirm nodes. */
3586 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b)
3588 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
3592 * Set the default node attribute compare operation for an ir_op_ops.
3594 * @param code the opcode for the default operation
3595 * @param ops the operations initialized
3600 static ir_op_ops *firm_set_default_node_cmp_attr(opcode code, ir_op_ops *ops)
3604 ops->node_cmp_attr = node_cmp_attr_##a; \
3631 * Compare function for two nodes in the hash table. Gets two
3632 * nodes as parameters. Returns 0 if the nodes are a cse.
3634 int identities_cmp(const void *elt, const void *key)
3642 if (a == b) return 0;
3644 if ((get_irn_op(a) != get_irn_op(b)) ||
3645 (get_irn_mode(a) != get_irn_mode(b))) return 1;
3647 /* compare if a's in and b's in are of equal length */
3648 irn_arity_a = get_irn_intra_arity (a);
3649 if (irn_arity_a != get_irn_intra_arity(b))
3652 /* for block-local cse and op_pin_state_pinned nodes: */
3653 if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) {
3654 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
3658 /* compare a->in[0..ins] with b->in[0..ins] */
3659 for (i = 0; i < irn_arity_a; i++)
3660 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
3664 * here, we already now that the nodes are identical except their
3667 if (a->op->ops.node_cmp_attr)
3668 return a->op->ops.node_cmp_attr(a, b);
3674 * Calculate a hash value of a node.
3676 unsigned ir_node_hash(ir_node *node)
3681 if (node->op == op_Const) {
3682 /* special value for const, as they only differ in their tarval. */
3683 h = HASH_PTR(node->attr.con.tv);
3684 h = 9*h + HASH_PTR(get_irn_mode(node));
3685 } else if (node->op == op_SymConst) {
3686 /* special value for const, as they only differ in their symbol. */
3687 h = HASH_PTR(node->attr.symc.sym.type_p);
3688 h = 9*h + HASH_PTR(get_irn_mode(node));
3691 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
3692 h = irn_arity = get_irn_intra_arity(node);
3694 /* consider all in nodes... except the block if not a control flow. */
3695 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
3696 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
3700 h = 9*h + HASH_PTR(get_irn_mode(node));
3702 h = 9*h + HASH_PTR(get_irn_op(node));
3708 pset *new_identities(void) {
3709 return new_pset(identities_cmp, N_IR_NODES);
3712 void del_identities(pset *value_table) {
3713 del_pset(value_table);
3717 * Return the canonical node computing the same value as n.
3718 * Looks up the node in a hash table.
3720 * For Const nodes this is performed in the constructor, too. Const
3721 * nodes are extremely time critical because of their frequent use in
3722 * constant string arrays.
3724 static INLINE ir_node *identify(pset *value_table, ir_node *n)
3728 if (!value_table) return n;
3730 if (get_opt_reassociation()) {
3731 if (is_op_commutative(get_irn_op(n))) {
3732 ir_node *l = get_binop_left(n);
3733 ir_node *r = get_binop_right(n);
3735 /* for commutative operators perform a OP b == b OP a */
3736 if (get_irn_idx(l) > get_irn_idx(r)) {
3737 set_binop_left(n, r);
3738 set_binop_right(n, l);
3743 o = pset_find(value_table, n, ir_node_hash(n));
3752 * During construction we set the op_pin_state_pinned flag in the graph right when the
3753 * optimization is performed. The flag turning on procedure global cse could
3754 * be changed between two allocations. This way we are safe.
3756 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
3759 n = identify(value_table, n);
3760 if (get_irn_n(old, -1) != get_irn_n(n, -1))
3761 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3766 * Return the canonical node computing the same value as n.
3767 * Looks up the node in a hash table, enters it in the table
3768 * if it isn't there yet.
3770 ir_node *identify_remember(pset *value_table, ir_node *n)
3774 if (!value_table) return n;
3776 if (get_opt_reassociation()) {
3777 if (is_op_commutative(get_irn_op(n))) {
3778 ir_node *l = get_binop_left(n);
3779 ir_node *r = get_binop_right(n);
3781 /* for commutative operators perform a OP b == b OP a */
3783 set_binop_left(n, r);
3784 set_binop_right(n, l);
3789 /* lookup or insert in hash table with given hash key. */
3790 o = pset_insert (value_table, n, ir_node_hash (n));
3799 /* Add a node to the identities value table. */
3800 void add_identities(pset *value_table, ir_node *node) {
3801 if (get_opt_cse() && is_no_Block(node))
3802 identify_remember(value_table, node);
3805 /* Visit each node in the value table of a graph. */
3806 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
3808 ir_graph *rem = current_ir_graph;
3810 current_ir_graph = irg;
3811 foreach_pset(irg->value_table, node)
3813 current_ir_graph = rem;
3817 * Garbage in, garbage out. If a node has a dead input, i.e., the
3818 * Bad node is input to the node, return the Bad node.
3820 static INLINE ir_node *gigo(ir_node *node)
3823 ir_op *op = get_irn_op(node);
3825 /* remove garbage blocks by looking at control flow that leaves the block
3826 and replacing the control flow by Bad. */
3827 if (get_irn_mode(node) == mode_X) {
3828 ir_node *block = get_nodes_block(skip_Proj(node));
3830 /* Don't optimize nodes in immature blocks. */
3831 if (!get_Block_matured(block)) return node;
3832 /* Don't optimize End, may have Bads. */
3833 if (op == op_End) return node;
3835 if (is_Block(block)) {
3836 irn_arity = get_irn_arity(block);
3837 for (i = 0; i < irn_arity; i++) {
3838 if (!is_Bad(get_irn_n(block, i)))
3841 if (i == irn_arity) return new_Bad();
3845 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
3846 blocks predecessors is dead. */
3847 if (op != op_Block && op != op_Phi && op != op_Tuple) {
3848 irn_arity = get_irn_arity(node);
3851 * Beware: we can only read the block of a non-floating node.
3853 if (is_irn_pinned_in_irg(node) &&
3854 is_Block_dead(get_nodes_block(node)))
3857 for (i = 0; i < irn_arity; i++) {
3858 ir_node *pred = get_irn_n(node, i);
3863 /* Propagating Unknowns here seems to be a bad idea, because
3864 sometimes we need a node as a input and did not want that
3866 However, it might be useful to move this into a later phase
3867 (if you think that optimizing such code is useful). */
3868 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
3869 return new_Unknown(get_irn_mode(node));
3874 /* With this code we violate the agreement that local_optimize
3875 only leaves Bads in Block, Phi and Tuple nodes. */
3876 /* If Block has only Bads as predecessors it's garbage. */
3877 /* If Phi has only Bads as predecessors it's garbage. */
3878 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
3879 irn_arity = get_irn_arity(node);
3880 for (i = 0; i < irn_arity; i++) {
3881 if (!is_Bad(get_irn_n(node, i))) break;
3883 if (i == irn_arity) node = new_Bad();
3890 * These optimizations deallocate nodes from the obstack.
3891 * It can only be called if it is guaranteed that no other nodes
3892 * reference this one, i.e., right after construction of a node.
3894 * current_ir_graph must be set to the graph of the node!
3896 ir_node *optimize_node(ir_node *n)
3900 opcode iro = get_irn_opcode(n);
3902 /* Always optimize Phi nodes: part of the construction. */
3903 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
3905 /* constant expression evaluation / constant folding */
3906 if (get_opt_constant_folding()) {
3907 /* neither constants nor Tuple values can be evaluated */
3908 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
3909 /* try to evaluate */
3910 tv = computed_value(n);
3911 if (tv != tarval_bad) {
3913 ir_type *old_tp = get_irn_type(n);
3914 int i, arity = get_irn_arity(n);
3918 * Try to recover the type of the new expression.
3920 for (i = 0; i < arity && !old_tp; ++i)
3921 old_tp = get_irn_type(get_irn_n(n, i));
3924 * we MUST copy the node here temporary, because it's still needed
3925 * for DBG_OPT_CSTEVAL
3927 node_size = offsetof(ir_node, attr) + n->op->attr_size;
3928 oldn = alloca(node_size);
3930 memcpy(oldn, n, node_size);
3931 CLONE_ARR_A(ir_node *, oldn->in, n->in);
3933 /* ARG, copy the in array, we need it for statistics */
3934 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
3936 /* note the inplace edges module */
3937 edges_node_deleted(n, current_ir_graph);
3939 /* evaluation was successful -- replace the node. */
3940 irg_kill_node(current_ir_graph, n);
3941 nw = new_Const(get_tarval_mode (tv), tv);
3943 if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv))
3944 set_Const_type(nw, old_tp);
3945 DBG_OPT_CSTEVAL(oldn, nw);
3951 /* remove unnecessary nodes */
3952 if (get_opt_constant_folding() ||
3953 (iro == iro_Phi) || /* always optimize these nodes. */
3955 (iro == iro_Proj) ||
3956 (iro == iro_Block) ) /* Flags tested local. */
3957 n = equivalent_node (n);
3959 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
3961 /* Common Subexpression Elimination.
3963 * Checks whether n is already available.
3964 * The block input is used to distinguish different subexpressions. Right
3965 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
3966 * subexpressions within a block.
3969 n = identify_cons (current_ir_graph->value_table, n);
3972 edges_node_deleted(oldn, current_ir_graph);
3974 /* We found an existing, better node, so we can deallocate the old node. */
3975 irg_kill_node(current_ir_graph, oldn);
3979 /* Some more constant expression evaluation that does not allow to
3981 iro = get_irn_opcode(n);
3982 if (get_opt_constant_folding() ||
3983 (iro == iro_Cond) ||
3984 (iro == iro_Proj) ||
3985 (iro == iro_Sel)) /* Flags tested local. */
3986 n = transform_node (n);
3988 /* Remove nodes with dead (Bad) input.
3989 Run always for transformation induced Bads. */
3992 /* Now we have a legal, useful node. Enter it in hash table for CSE */
3993 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
3994 n = identify_remember (current_ir_graph->value_table, n);
4002 * These optimizations never deallocate nodes (in place). This can cause dead
4003 * nodes lying on the obstack. Remove these by a dead node elimination,
4004 * i.e., a copying garbage collection.
4006 ir_node *optimize_in_place_2(ir_node *n)
4010 opcode iro = get_irn_opcode(n);
4012 if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n;
4014 /* constant expression evaluation / constant folding */
4015 if (get_opt_constant_folding()) {
4016 /* neither constants nor Tuple values can be evaluated */
4017 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
4018 /* try to evaluate */
4019 tv = computed_value(n);
4020 if (tv != tarval_bad) {
4021 /* evaluation was successful -- replace the node. */
4022 ir_type *old_tp = get_irn_type(n);
4023 int i, arity = get_irn_arity(n);
4026 * Try to recover the type of the new expression.
4028 for (i = 0; i < arity && !old_tp; ++i)
4029 old_tp = get_irn_type(get_irn_n(n, i));
4031 n = new_Const(get_tarval_mode(tv), tv);
4033 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
4034 set_Const_type(n, old_tp);
4036 DBG_OPT_CSTEVAL(oldn, n);
4042 /* remove unnecessary nodes */
4043 if (get_opt_constant_folding() ||
4044 (iro == iro_Phi) || /* always optimize these nodes. */
4045 (iro == iro_Id) || /* ... */
4046 (iro == iro_Proj) || /* ... */
4047 (iro == iro_Block) ) /* Flags tested local. */
4048 n = equivalent_node(n);
4050 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
4052 /** common subexpression elimination **/
4053 /* Checks whether n is already available. */
4054 /* The block input is used to distinguish different subexpressions. Right
4055 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
4056 subexpressions within a block. */
4057 if (get_opt_cse()) {
4058 n = identify(current_ir_graph->value_table, n);
4061 /* Some more constant expression evaluation. */
4062 iro = get_irn_opcode(n);
4063 if (get_opt_constant_folding() ||
4064 (iro == iro_Cond) ||
4065 (iro == iro_Proj) ||
4066 (iro == iro_Sel)) /* Flags tested local. */
4067 n = transform_node(n);
4069 /* Remove nodes with dead (Bad) input.
4070 Run always for transformation induced Bads. */
4073 /* Now we can verify the node, as it has no dead inputs any more. */
4076 /* Now we have a legal, useful node. Enter it in hash table for cse.
4077 Blocks should be unique anyways. (Except the successor of start:
4078 is cse with the start block!) */
4079 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
4080 n = identify_remember(current_ir_graph->value_table, n);
4086 * Wrapper for external use, set proper status bits after optimization.
4088 ir_node *optimize_in_place(ir_node *n)
4090 /* Handle graph state */
4091 assert(get_irg_phase_state(current_ir_graph) != phase_building);
4093 if (get_opt_global_cse())
4094 set_irg_pinned(current_ir_graph, op_pin_state_floats);
4095 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
4096 set_irg_outs_inconsistent(current_ir_graph);
4098 /* FIXME: Maybe we could also test whether optimizing the node can
4099 change the control graph. */
4100 set_irg_doms_inconsistent(current_ir_graph);
4101 return optimize_in_place_2 (n);
4105 * Sets the default operation for an ir_ops.
4107 ir_op_ops *firm_set_default_operations(opcode code, ir_op_ops *ops)
4109 ops = firm_set_default_computed_value(code, ops);
4110 ops = firm_set_default_equivalent_node(code, ops);
4111 ops = firm_set_default_transform_node(code, ops);
4112 ops = firm_set_default_node_cmp_attr(code, ops);
4113 ops = firm_set_default_get_type(code, ops);
4114 ops = firm_set_default_get_type_attr(code, ops);
4115 ops = firm_set_default_get_entity_attr(code, ops);