3 * File name: ir/ir/iropt.c
4 * Purpose: iropt --- optimizations intertwined with IR construction.
5 * Author: Christian Schaefer
6 * Modified by: Goetz Lindenmaier
9 * Copyright: (c) 1998-2003 Universität Karlsruhe
10 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
17 # include "irnode_t.h"
18 # include "irgraph_t.h"
24 # include "dbginfo_t.h"
25 # include "iropt_dbg.h"
26 # include "irflag_t.h"
28 /* Make types visible to allow most efficient access */
29 # include "entity_t.h"
32 * Trivial INLINEable routine for copy propagation.
33 * Does follow Ids, needed to optimize INLINEd code.
35 static INLINE ir_node *
36 follow_Id (ir_node *n)
38 while (get_irn_op (n) == op_Id) n = get_Id_pred (n);
43 * Returns the tarval of a Const node or tarval_bad for all other nodes.
45 static INLINE tarval *
48 if ((n != NULL) && (get_irn_op(n) == op_Const))
49 return get_Const_tarval(n); /* might return tarval_bad */
54 static tarval *computed_value_Const(ir_node *n)
56 return get_Const_tarval(n);
59 static tarval *computed_value_SymConst(ir_node *n)
61 if ((get_SymConst_kind(n) == size) &&
62 (get_type_state(get_SymConst_type(n))) == layout_fixed)
63 return new_tarval_from_long (get_type_size(get_SymConst_type(n)), mode_Is);
67 static tarval *computed_value_Add(ir_node *n)
69 ir_node *a = get_Add_left(n);
70 ir_node *b = get_Add_right(n);
72 tarval *ta = value_of(a);
73 tarval *tb = value_of(b);
75 if ((ta != tarval_bad) && (tb != tarval_bad)
76 && (get_irn_mode(a) == get_irn_mode(b))
77 && !(get_mode_sort(get_irn_mode(a)) == irms_reference)) {
78 return tarval_add(ta, tb);
83 static tarval *computed_value_Sub(ir_node *n)
85 ir_node *a = get_Sub_left(n);
86 ir_node *b = get_Sub_right(n);
88 tarval *ta = value_of(a);
89 tarval *tb = value_of(b);
91 if ((ta != tarval_bad) && (tb != tarval_bad)
92 && (get_irn_mode(a) == get_irn_mode(b))
93 && !(get_mode_sort(get_irn_mode(a)) == irms_reference)) {
94 return tarval_sub(ta, tb);
99 static tarval *computed_value_Minus(ir_node *n)
101 ir_node *a = get_Minus_op(n);
102 tarval *ta = value_of(a);
104 if ((ta != tarval_bad) && mode_is_signed(get_irn_mode(a)))
105 return tarval_neg(ta);
110 static tarval *computed_value_Mul(ir_node *n)
112 ir_node *a = get_Mul_left(n);
113 ir_node *b = get_Mul_right(n);
115 tarval *ta = value_of(a);
116 tarval *tb = value_of(b);
118 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
119 return tarval_mul(ta, tb);
121 /* a*0 = 0 or 0*b = 0:
122 calls computed_value recursive and returns the 0 with proper
126 if ( ( ((v = ta) != tarval_bad)
127 && (v == get_mode_null(get_tarval_mode(v))) )
128 || ( ((v = tb) != tarval_bad)
129 && (v == get_mode_null(get_tarval_mode(v))) )) {
136 static tarval *computed_value_Quot(ir_node *n)
138 ir_node *a = get_Quot_left(n);
139 ir_node *b = get_Quot_right(n);
141 tarval *ta = value_of(a);
142 tarval *tb = value_of(b);
144 /* This was missing in original implementation. Why? */
145 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
146 if (tb != get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */
147 return tarval_quo(ta, tb);
152 static tarval *computed_value_Div(ir_node *n)
154 ir_node *a = get_Div_left(n);
155 ir_node *b = get_Div_right(n);
157 tarval *ta = value_of(a);
158 tarval *tb = value_of(b);
160 /* This was missing in original implementation. Why? */
161 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
162 if (tb != get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */
163 return tarval_div(ta, tb);
168 static tarval *computed_value_Mod(ir_node *n)
170 ir_node *a = get_Mod_left(n);
171 ir_node *b = get_Mod_right(n);
173 tarval *ta = value_of(a);
174 tarval *tb = value_of(b);
176 /* This was missing in original implementation. Why? */
177 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
178 if (tb != get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */
179 return tarval_mod(ta, tb);
184 static tarval *computed_value_Abs(ir_node *n)
186 ir_node *a = get_Abs_op(n);
187 tarval *ta = value_of(a);
189 if (ta != tarval_bad)
190 return tarval_abs(ta);
195 static tarval *computed_value_And(ir_node *n)
197 ir_node *a = get_And_left(n);
198 ir_node *b = get_And_right(n);
200 tarval *ta = value_of(a);
201 tarval *tb = value_of(b);
203 if ((ta != tarval_bad) && (tb != tarval_bad)) {
204 return tarval_and (ta, tb);
208 if ( (tarval_classify ((v = computed_value (a))) == TV_CLASSIFY_NULL)
209 || (tarval_classify ((v = computed_value (b))) == TV_CLASSIFY_NULL)) {
216 static tarval *computed_value_Or(ir_node *n)
218 ir_node *a = get_Or_left(n);
219 ir_node *b = get_Or_right(n);
221 tarval *ta = value_of(a);
222 tarval *tb = value_of(b);
224 if ((ta != tarval_bad) && (tb != tarval_bad)) {
225 return tarval_or (ta, tb);
228 if ( (tarval_classify ((v = computed_value (a))) == TV_CLASSIFY_ALL_ONE)
229 || (tarval_classify ((v = computed_value (b))) == TV_CLASSIFY_ALL_ONE)) {
236 static tarval *computed_value_Eor(ir_node *n)
238 ir_node *a = get_Eor_left(n);
239 ir_node *b = get_Eor_right(n);
241 tarval *ta = value_of(a);
242 tarval *tb = value_of(b);
244 if ((ta != tarval_bad) && (tb != tarval_bad)) {
245 return tarval_eor (ta, tb);
250 static tarval *computed_value_Not(ir_node *n)
252 ir_node *a = get_Not_op(n);
253 tarval *ta = value_of(a);
255 if (ta != tarval_bad)
256 return tarval_not(ta);
261 static tarval *computed_value_Shl(ir_node *n)
263 ir_node *a = get_Shl_left(n);
264 ir_node *b = get_Shl_right(n);
266 tarval *ta = value_of(a);
267 tarval *tb = value_of(b);
269 if ((ta != tarval_bad) && (tb != tarval_bad)) {
270 return tarval_shl (ta, tb);
275 static tarval *computed_value_Shr(ir_node *n)
277 ir_node *a = get_Shr_left(n);
278 ir_node *b = get_Shr_right(n);
280 tarval *ta = value_of(a);
281 tarval *tb = value_of(b);
283 if ((ta != tarval_bad) && (tb != tarval_bad)) {
284 return tarval_shr (ta, tb);
289 static tarval *computed_value_Shrs(ir_node *n)
291 ir_node *a = get_Shrs_left(n);
292 ir_node *b = get_Shrs_right(n);
294 tarval *ta = value_of(a);
295 tarval *tb = value_of(b);
297 if ((ta != tarval_bad) && (tb != tarval_bad)) {
298 return tarval_shrs (ta, tb);
303 static tarval *computed_value_Rot(ir_node *n)
305 ir_node *a = get_Rot_left(n);
306 ir_node *b = get_Rot_right(n);
308 tarval *ta = value_of(a);
309 tarval *tb = value_of(b);
311 if ((ta != tarval_bad) && (tb != tarval_bad)) {
312 /* return tarval_rot (ta, tb); */
317 static tarval *computed_value_Conv(ir_node *n)
319 ir_node *a = get_Conv_op(n);
320 tarval *ta = value_of(a);
322 if (ta != tarval_bad)
323 return tarval_convert_to(ta, get_irn_mode(n));
328 static tarval *computed_value_Proj(ir_node *n)
330 ir_node *a = get_Proj_pred(n), *b;
333 /* Optimize Cmp nodes.
334 This performs a first step of unreachable code elimination.
335 Proj can not be computed, but folding a Cmp above the Proj here is
336 not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
338 There are several case where we can evaluate a Cmp node:
339 1. The nodes compared are both the same. If we compare for
340 equal, greater equal, ... this will return true, else it
341 will return false. This step relies on cse.
342 2. The predecessors of Cmp are target values. We can evaluate
344 3. The predecessors are Allocs or void* constants. Allocs never
345 return NULL, they raise an exception. Therefore we can predict
347 if (get_irn_op(a) == op_Cmp) {
348 aa = get_Cmp_left(a);
349 ab = get_Cmp_right(a);
351 if (aa == ab) { /* 1.: */
352 /* This is a tric with the bits used for encoding the Cmp
353 Proj numbers, the following statement is not the same:
354 return new_tarval_from_long ((get_Proj_proj(n) == Eq), mode_b) */
355 return new_tarval_from_long ((get_Proj_proj(n) & Eq), mode_b);
357 tarval *taa = computed_value (aa);
358 tarval *tab = computed_value (ab);
360 if ((taa != tarval_bad) && (tab != tarval_bad)) { /* 2.: */
361 /* strange checks... */
362 pnc_number flags = tarval_cmp (taa, tab);
363 if (flags != False) {
364 return new_tarval_from_long (get_Proj_proj(n) & flags, mode_b);
366 } else { /* check for 3.: */
367 ir_node *aaa = skip_nop(skip_Proj(aa));
368 ir_node *aba = skip_nop(skip_Proj(ab));
370 if ( ( (/* aa is ProjP and aaa is Alloc */
371 (get_irn_op(aa) == op_Proj)
372 && (mode_is_reference(get_irn_mode(aa)))
373 && (get_irn_op(aaa) == op_Alloc))
374 && ( (/* ab is constant void */
375 (get_irn_op(ab) == op_Const)
376 && (mode_is_reference(get_irn_mode(ab)))
377 && (get_Const_tarval(ab) == get_mode_null(get_irn_mode(ab))))
378 || (/* ab is other Alloc */
379 (get_irn_op(ab) == op_Proj)
380 && (mode_is_reference(get_irn_mode(ab)))
381 && (get_irn_op(aba) == op_Alloc)
383 || (/* aa is void and aba is Alloc */
384 (get_irn_op(aa) == op_Const)
385 && (mode_is_reference(get_irn_mode(aa)))
386 && (get_Const_tarval(aa) == get_mode_null(get_irn_mode(aa)))
387 && (get_irn_op(ab) == op_Proj)
388 && (mode_is_reference(get_irn_mode(ab)))
389 && (get_irn_op(aba) == op_Alloc)))
391 return new_tarval_from_long (get_Proj_proj(n) & Ne, mode_b);
394 } else if (get_irn_op(a) == op_DivMod) {
395 tarval *tb = value_of(b = get_DivMod_right(a));
396 tarval *ta = value_of(a = get_DivMod_left(a));
398 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
399 if (tb == get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */
401 if (get_Proj_proj(n)== 0) /* Div */
402 return tarval_div(ta, tb);
404 return tarval_mod(ta, tb);
411 * If the parameter n can be computed, return its value, else tarval_bad.
412 * Performs constant folding.
414 * GL: Only if n is arithmetic operator?
416 tarval *computed_value(ir_node *n)
418 if (n->op->computed_value)
419 return n->op->computed_value(n);
424 * set the default computed_value evaluator
426 static ir_op *firm_set_default_computed_value(ir_op *op)
430 op->computed_value = computed_value_##a; \
455 op->computed_value = NULL;
463 /* returns 1 if the a and b are pointers to different locations. */
465 different_identity (ir_node *a, ir_node *b)
467 assert (mode_is_reference(get_irn_mode (a))
468 && mode_is_reference(get_irn_mode (b)));
470 if (get_irn_op (a) == op_Proj && get_irn_op(b) == op_Proj) {
471 ir_node *a1 = get_Proj_pred (a);
472 ir_node *b1 = get_Proj_pred (b);
473 if (a1 != b1 && get_irn_op (a1) == op_Alloc
474 && get_irn_op (b1) == op_Alloc)
481 static ir_node *equivalent_node_Block(ir_node *n)
485 /* The Block constructor does not call optimize, but mature_block
486 calls the optimization. */
487 assert(get_Block_matured(n));
489 /* Straightening: a single entry Block following a single exit Block
490 can be merged, if it is not the Start block. */
491 /* !!! Beware, all Phi-nodes of n must have been optimized away.
492 This should be true, as the block is matured before optimize is called.
493 But what about Phi-cycles with the Phi0/Id that could not be resolved?
494 Remaining Phi nodes are just Ids. */
495 if ((get_Block_n_cfgpreds(n) == 1) &&
496 (get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp) &&
497 (get_opt_control_flow_straightening())) {
498 n = get_nodes_Block(get_Block_cfgpred(n, 0)); DBG_OPT_STG;
500 } else if ((get_Block_n_cfgpreds(n) == 2) &&
501 (get_opt_control_flow_weak_simplification())) {
502 /* Test whether Cond jumps twice to this block
503 @@@ we could do this also with two loops finding two preds from several ones. */
504 ir_node *a = get_Block_cfgpred(n, 0);
505 ir_node *b = get_Block_cfgpred(n, 1);
507 if ((get_irn_op(a) == op_Proj) &&
508 (get_irn_op(b) == op_Proj) &&
509 (get_Proj_pred(a) == get_Proj_pred(b)) &&
510 (get_irn_op(get_Proj_pred(a)) == op_Cond) &&
511 (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) {
512 /* Also a single entry Block following a single exit Block. Phis have
513 twice the same operand and will be optimized away. */
514 n = get_nodes_Block(a); DBG_OPT_IFSIM;
516 } else if (get_opt_unreachable_code() &&
517 (n != current_ir_graph->start_block) &&
518 (n != current_ir_graph->end_block) ) {
520 /* If all inputs are dead, this block is dead too, except if it is
521 the start or end block. This is a step of unreachable code
523 for (i = 0; i < get_Block_n_cfgpreds(n); i++) {
524 if (!is_Bad(get_Block_cfgpred(n, i))) break;
526 if (i == get_Block_n_cfgpreds(n))
533 static ir_node *equivalent_node_Jmp(ir_node *n)
535 /* GL: Why not same for op_Raise?? */
536 /* unreachable code elimination */
537 if (is_Bad(get_nodes_Block(n)))
543 static ir_node *equivalent_node_Cond(ir_node *n)
545 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
546 See cases for iro_Cond and iro_Proj in transform_node. */
550 static ir_node *equivalent_node_Or(ir_node *n)
552 ir_node *a = get_Or_left(n);
553 ir_node *b = get_Or_right(n);
562 static ir_node *equivalent_node_neutral_zero(ir_node *n)
566 ir_node *a = get_binop_left(n);
567 ir_node *b = get_binop_right(n);
572 /* After running compute_node there is only one constant predecessor.
573 Find this predecessors value and remember the other node: */
574 if ((tv = computed_value (a)) != tarval_bad) {
576 } else if ((tv = computed_value (b)) != tarval_bad) {
581 /* If this predecessors constant value is zero, the operation is
582 unnecessary. Remove it: */
583 if (tarval_classify (tv) == TV_CLASSIFY_NULL) {
584 n = on; DBG_OPT_ALGSIM1;
590 static ir_node *equivalent_node_Add(ir_node *n)
592 return equivalent_node_neutral_zero(n);
595 static ir_node *equivalent_node_Eor(ir_node *n)
597 return equivalent_node_neutral_zero(n);
600 static ir_node *equivalent_node_left_zero(ir_node *n)
604 ir_node *a = get_binop_left(n);
605 ir_node *b = get_binop_right(n);
607 /* optimize operations that are not commutative but have neutral 0 on left. Test only one predecessor. */
608 if (tarval_classify (computed_value (b)) == TV_CLASSIFY_NULL) {
609 n = a; DBG_OPT_ALGSIM1;
615 static ir_node *equivalent_node_Sub(ir_node *n)
617 return equivalent_node_left_zero(n);
620 static ir_node *equivalent_node_Shl(ir_node *n)
622 return equivalent_node_left_zero(n);
625 static ir_node *equivalent_node_Shr(ir_node *n)
627 return equivalent_node_left_zero(n);
630 static ir_node *equivalent_node_Shrs(ir_node *n)
632 return equivalent_node_left_zero(n);
635 static ir_node *equivalent_node_Rot(ir_node *n)
637 return equivalent_node_left_zero(n);
640 static ir_node *equivalent_node_symmetric_unop(ir_node *n)
644 /* optimize symmetric unop */
645 if (get_irn_op(get_unop_op(n)) == get_irn_op(n)) {
646 n = get_unop_op(get_unop_op(n)); DBG_OPT_ALGSIM2;
651 static ir_node *equivalent_node_Not(ir_node *n)
654 return equivalent_node_symmetric_unop(n);
657 static ir_node *equivalent_node_Minus(ir_node *n)
659 /* --x == x */ /* ??? Is this possible or can --x raise an
660 out of bounds exception if min =! max? */
661 return equivalent_node_symmetric_unop(n);
664 static ir_node *equivalent_node_Mul(ir_node *n)
668 ir_node *a = get_Mul_left(n);
669 ir_node *b = get_Mul_right(n);
671 /* Mul is commutative and has again an other neutral element. */
672 if (tarval_classify (computed_value (a)) == TV_CLASSIFY_ONE) {
673 n = b; DBG_OPT_ALGSIM1;
674 } else if (tarval_classify (computed_value (b)) == TV_CLASSIFY_ONE) {
675 n = a; DBG_OPT_ALGSIM1;
680 static ir_node *equivalent_node_Div(ir_node *n)
682 ir_node *a = get_Div_left(n);
683 ir_node *b = get_Div_right(n);
685 /* Div is not commutative. */
686 if (tarval_classify (computed_value (b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
687 /* Turn Div into a tuple (mem, bad, a) */
688 ir_node *mem = get_Div_mem(n);
689 turn_into_tuple(n, 3);
690 set_Tuple_pred(n, 0, mem);
691 set_Tuple_pred(n, 1, new_Bad());
692 set_Tuple_pred(n, 2, a);
697 static ir_node *equivalent_node_And(ir_node *n)
701 ir_node *a = get_And_left(n);
702 ir_node *b = get_And_right(n);
705 n = a; /* And has it's own neutral element */
706 } else if (tarval_classify (computed_value (a)) == TV_CLASSIFY_ALL_ONE) {
708 } else if (tarval_classify (computed_value (b)) == TV_CLASSIFY_ALL_ONE) {
711 if (n != oldn) DBG_OPT_ALGSIM1;
715 static ir_node *equivalent_node_Conv(ir_node *n)
718 ir_node *a = get_Conv_op(n);
721 ir_mode *n_mode = get_irn_mode(n);
722 ir_mode *a_mode = get_irn_mode(a);
724 if (n_mode == a_mode) { /* No Conv necessary */
725 n = a; DBG_OPT_ALGSIM3;
726 } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */
730 n_mode = get_irn_mode(n);
731 b_mode = get_irn_mode(b);
733 if (n_mode == b_mode) {
734 if (n_mode == mode_b) {
735 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */ DBG_OPT_ALGSIM1;
737 else if (mode_is_int(n_mode) || mode_is_character(n_mode)) {
738 if (smaller_mode(b_mode, a_mode)){
739 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */ DBG_OPT_ALGSIM1;
747 static ir_node *equivalent_node_Phi(ir_node *n)
749 /* Several optimizations:
750 - no Phi in start block.
751 - remove Id operators that are inputs to Phi
752 - fold Phi-nodes, iff they have only one predecessor except
758 ir_node *block = NULL; /* to shutup gcc */
759 ir_node *first_val = NULL; /* to shutup gcc */
760 ir_node *scnd_val = NULL; /* to shutup gcc */
762 if (!get_opt_normalize()) return n;
764 n_preds = get_Phi_n_preds(n);
766 block = get_nodes_Block(n);
767 /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
768 assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
769 if ((is_Bad(block)) || /* Control dead */
770 (block == current_ir_graph->start_block)) /* There should be no Phi nodes */
771 return new_Bad(); /* in the Start Block. */
773 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
776 /* first we test for a special case: */
777 /* Confirm is a special node fixing additional information for a
778 value that is known at a certain point. This is useful for
779 dataflow analysis. */
781 ir_node *a = follow_Id (get_Phi_pred(n, 0));
782 ir_node *b = follow_Id (get_Phi_pred(n, 1));
783 if ( (get_irn_op(a) == op_Confirm)
784 && (get_irn_op(b) == op_Confirm)
785 && (follow_Id (get_irn_n(a, 0)) == follow_Id(get_irn_n(b, 0)))
786 && (get_irn_n(a, 1) == get_irn_n (b, 1))
787 && (a->data.num == (~b->data.num & irpn_True) )) {
788 return follow_Id (get_irn_n(a, 0));
793 /* Find first non-self-referencing input */
794 for (i = 0; i < n_preds; ++i) {
795 first_val = follow_Id(get_Phi_pred(n, i));
797 set_Phi_pred(n, i, first_val);
798 if ( (first_val != n) /* not self pointer */
799 && (get_irn_op(first_val) != op_Bad) /* value not dead */
800 && !(is_Bad (get_Block_cfgpred(block, i))) ) { /* not dead control flow */
801 break; /* then found first value. */
805 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
806 if (i >= n_preds) { return new_Bad(); }
810 /* follow_Id () for rest of inputs, determine if any of these
811 are non-self-referencing */
812 while (++i < n_preds) {
813 scnd_val = follow_Id(get_Phi_pred(n, i));
815 set_Phi_pred(n, i, scnd_val);
817 && (scnd_val != first_val)
818 && (get_irn_op(scnd_val) != op_Bad)
819 && !(is_Bad (get_Block_cfgpred(block, i))) ) {
824 /* Fold, if no multiple distinct non-self-referencing inputs */
826 n = first_val; DBG_OPT_PHI;
828 /* skip the remaining Ids. */
829 while (++i < n_preds) {
830 set_Phi_pred(n, i, follow_Id(get_Phi_pred(n, i)));
836 static ir_node *equivalent_node_Load(ir_node *n)
838 #if 0 /* Is an illegal transformation: different nodes can
839 represent the same pointer value!! */
840 ir_node *a = skip_Proj(get_Load_mem(n));
841 ir_node *b = get_Load_ptr(n);
843 if (get_irn_op(a) == op_Store) {
844 if ( different_identity (b, get_Store_ptr(a))) {
845 /* load and store use different pointers, therefore load
846 needs not take store's memory but the state before. */
847 set_Load_mem (n, get_Store_mem(a));
848 } else if (( 0 /* ???didn't get cryptic test that returns 0 */ )) {
855 static ir_node *equivalent_node_Store(ir_node *n)
859 /* remove unnecessary store. */
860 ir_node *a = skip_Proj(get_Store_mem(n));
861 ir_node *b = get_Store_ptr(n);
862 ir_node *c = skip_Proj(get_Store_value(n));
864 if (get_irn_op(a) == op_Store
865 && get_Store_ptr(a) == b
866 && skip_Proj(get_Store_value(a)) == c) {
867 /* We have twice exactly the same store -- a write after write. */
869 } else if (get_irn_op(c) == op_Load
870 && (a == c || skip_Proj(get_Load_mem(c)) == a)
871 && get_Load_ptr(c) == b ) {
872 /* We just loaded the value from the same memory, i.e., the store
873 doesn't change the memory -- a write after read. */
874 a = get_Store_mem(n);
875 turn_into_tuple(n, 2);
876 set_Tuple_pred(n, 0, a);
877 set_Tuple_pred(n, 1, new_Bad()); DBG_OPT_WAR;
882 static ir_node *equivalent_node_Proj(ir_node *n)
886 ir_node *a = get_Proj_pred(n);
888 if ( get_irn_op(a) == op_Tuple) {
889 /* Remove the Tuple/Proj combination. */
890 if ( get_Proj_proj(n) <= get_Tuple_n_preds(a) ) {
891 n = get_Tuple_pred(a, get_Proj_proj(n)); DBG_OPT_TUPLE;
893 assert(0); /* This should not happen! */
896 } else if (get_irn_mode(n) == mode_X &&
897 is_Bad(get_nodes_Block(n))) {
898 /* Remove dead control flow -- early gigo. */
904 static ir_node *equivalent_node_Id(ir_node *n)
908 n = follow_Id (n); DBG_OPT_ID;
913 case iro_Mod, Quot, DivMod
914 DivMod allocates new nodes --> it's treated in transform node.
915 What about Quot, DivMod?
919 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
920 * perform no actual computation, as, e.g., the Id nodes. It does not create
921 * new nodes. It is therefore safe to free n if the node returned is not n.
922 * If a node returns a Tuple we can not just skip it. If the size of the
923 * in array fits, we transform n into a tuple (e.g., Div).
926 equivalent_node (ir_node *n)
928 if (n->op->equivalent_node)
929 return n->op->equivalent_node(n);
934 * set the default equivalent node operation
936 static ir_op *firm_set_default_equivalent_node(ir_op *op)
940 op->equivalent_node = equivalent_node_##a; \
967 op->equivalent_node = NULL;
975 * Do node specific optimizations of nodes predecessors.
978 optimize_preds(ir_node *n) {
979 ir_node *a = NULL, *b = NULL;
981 /* get the operands we will work on for simple cases. */
983 a = get_binop_left(n);
984 b = get_binop_right(n);
985 } else if (is_unop(n)) {
989 switch (get_irn_opcode(n)) {
992 /* We don't want Cast as input to Cmp. */
993 if (get_irn_op(a) == op_Cast) {
997 if (get_irn_op(b) == op_Cast) {
1007 static ir_node *transform_node_Div(ir_node *n)
1009 tarval *ta = computed_value(n);
1011 if (ta != tarval_bad) {
1012 /* Turn Div into a tuple (mem, bad, value) */
1013 ir_node *mem = get_Div_mem(n);
1015 turn_into_tuple(n, 3);
1016 set_Tuple_pred(n, 0, mem);
1017 set_Tuple_pred(n, 1, new_Bad());
1018 set_Tuple_pred(n, 2, new_Const(get_tarval_mode(ta), ta));
1023 static ir_node *transform_node_Mod(ir_node *n)
1025 tarval *ta = computed_value(n);
1027 if (ta != tarval_bad) {
1028 /* Turn Mod into a tuple (mem, bad, value) */
1029 ir_node *mem = get_Mod_mem(n);
1030 turn_into_tuple(n, 3);
1031 set_Tuple_pred(n, 0, mem);
1032 set_Tuple_pred(n, 1, new_Bad());
1033 set_Tuple_pred(n, 2, new_Const(get_tarval_mode(ta), ta));
1038 static ir_node *transform_node_DivMod(ir_node *n)
1042 ir_node *a = get_DivMod_left(n);
1043 ir_node *b = get_DivMod_right(n);
1044 ir_mode *mode = get_irn_mode(a);
1046 if (!(mode_is_int(mode) && mode_is_int(get_irn_mode(b))))
1050 a = new_Const(mode, get_mode_one(mode));
1051 b = new_Const(mode, get_mode_null(mode));
1054 tarval *ta = value_of(a);
1055 tarval *tb = value_of(b);
1057 if (tb != tarval_bad) {
1058 if (tb == get_mode_one(get_tarval_mode(tb))) {
1059 b = new_Const (mode, get_mode_null(mode));
1061 } else if (ta != tarval_bad) {
1062 tarval *resa, *resb;
1063 resa = tarval_div (ta, tb);
1064 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
1065 Jmp for X result!? */
1066 resb = tarval_mod (ta, tb);
1067 if (resb == tarval_bad) return n; /* Causes exception! */
1068 a = new_Const (mode, resa);
1069 b = new_Const (mode, resb);
1072 } else if (ta == get_mode_null(get_tarval_mode(ta))) {
1077 if (evaluated) { /* replace by tuple */
1078 ir_node *mem = get_DivMod_mem(n);
1079 turn_into_tuple(n, 4);
1080 set_Tuple_pred(n, 0, mem);
1081 set_Tuple_pred(n, 1, new_Bad()); /* no exception */
1082 set_Tuple_pred(n, 2, a);
1083 set_Tuple_pred(n, 3, b);
1084 assert(get_nodes_Block(n));
1090 static ir_node *transform_node_Cond(ir_node *n)
1092 /* Replace the Cond by a Jmp if it branches on a constant
1095 ir_node *a = get_Cond_selector(n);
1096 tarval *ta = value_of(a);
1098 if ((ta != tarval_bad) &&
1099 (get_irn_mode(a) == mode_b) &&
1100 (get_opt_unreachable_code())) {
1101 /* It's a boolean Cond, branching on a boolean constant.
1102 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
1103 jmp = new_r_Jmp(current_ir_graph, get_nodes_Block(n));
1104 turn_into_tuple(n, 2);
1105 if (ta == tarval_b_true) {
1106 set_Tuple_pred(n, 0, new_Bad());
1107 set_Tuple_pred(n, 1, jmp);
1109 set_Tuple_pred(n, 0, jmp);
1110 set_Tuple_pred(n, 1, new_Bad());
1112 /* We might generate an endless loop, so keep it alive. */
1113 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_Block(n));
1114 } else if ((ta != tarval_bad) &&
1115 (get_irn_mode(a) == mode_Iu) &&
1116 (get_Cond_kind(n) == dense) &&
1117 (get_opt_unreachable_code())) {
1118 /* I don't want to allow Tuples smaller than the biggest Proj.
1119 Also this tuple might get really big...
1120 I generate the Jmp here, and remember it in link. Link is used
1121 when optimizing Proj. */
1122 set_irn_link(n, new_r_Jmp(current_ir_graph, get_nodes_Block(n)));
1123 /* We might generate an endless loop, so keep it alive. */
1124 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_Block(n));
1125 } else if ((get_irn_op(a) == op_Eor)
1126 && (get_irn_mode(a) == mode_b)
1127 && (tarval_classify(computed_value(get_Eor_right(a))) == TV_CLASSIFY_ONE)) {
1128 /* The Eor is a negate. Generate a new Cond without the negate,
1129 simulate the negate by exchanging the results. */
1130 set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_Block(n),
1132 } else if ((get_irn_op(a) == op_Not)
1133 && (get_irn_mode(a) == mode_b)) {
1134 /* A Not before the Cond. Generate a new Cond without the Not,
1135 simulate the Not by exchanging the results. */
1136 set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_Block(n),
1142 static ir_node *transform_node_Eor(ir_node *n)
1144 ir_node *a = get_Eor_left(n);
1145 ir_node *b = get_Eor_right(n);
1147 if ((get_irn_mode(n) == mode_b)
1148 && (get_irn_op(a) == op_Proj)
1149 && (get_irn_mode(a) == mode_b)
1150 && (tarval_classify (computed_value (b)) == TV_CLASSIFY_ONE)
1151 && (get_irn_op(get_Proj_pred(a)) == op_Cmp))
1152 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
1153 n = new_r_Proj(current_ir_graph, get_nodes_Block(n), get_Proj_pred(a),
1154 mode_b, get_negated_pnc(get_Proj_proj(a)));
1155 else if ((get_irn_mode(n) == mode_b)
1156 && (tarval_classify (computed_value (b)) == TV_CLASSIFY_ONE))
1157 /* The Eor is a Not. Replace it by a Not. */
1158 /* ????!!!Extend to bitfield 1111111. */
1159 n = new_r_Not(current_ir_graph, get_nodes_Block(n), a, mode_b);
1164 static ir_node *transform_node_Not(ir_node *n)
1166 ir_node *a = get_Not_op(n);
1168 if ( (get_irn_mode(n) == mode_b)
1169 && (get_irn_op(a) == op_Proj)
1170 && (get_irn_mode(a) == mode_b)
1171 && (get_irn_op(get_Proj_pred(a)) == op_Cmp))
1172 /* We negate a Cmp. The Cmp has the negated result anyways! */
1173 n = new_r_Proj(current_ir_graph, get_nodes_Block(n), get_Proj_pred(a),
1174 mode_b, get_negated_pnc(get_Proj_proj(a)));
1181 * Tries several [inplace] [optimizing] transformations and returns an
1182 * equivalent node. The difference to equivalent_node() is that these
1183 * transformations _do_ generate new nodes, and thus the old node must
1184 * not be freed even if the equivalent node isn't the old one.
1186 static ir_node *transform_node(ir_node *n)
1188 if (n->op->transform_node)
1189 n = n->op->transform_node(n);
1194 * set the default transform node operation
1196 static ir_op *firm_set_default_transform_node(ir_op *op)
1200 op->transform_node = transform_node_##a; \
1211 op->transform_node = NULL;
1219 /* **************** Common Subexpression Elimination **************** */
1221 /** The size of the hash table used, should estimate the number of nodes
1223 #define N_IR_NODES 512
1225 static int node_cmp_attr_Const(ir_node *a, ir_node *b)
1227 return (get_Const_tarval(a) != get_Const_tarval(b))
1228 || (get_Const_type(a) != get_Const_type(b));
1231 static int node_cmp_attr_Proj(ir_node *a, ir_node *b)
1233 return get_irn_proj_attr (a) != get_irn_proj_attr (b);
1236 static int node_cmp_attr_Filter(ir_node *a, ir_node *b)
1238 return get_Filter_proj(a) != get_Filter_proj(b);
1241 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b)
1243 return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where)
1244 || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type);
1247 static int node_cmp_attr_Free(ir_node *a, ir_node *b)
1249 return (get_irn_free_attr(a) != get_irn_free_attr(b));
1252 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b)
1254 return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num)
1255 || (get_irn_symconst_attr(a).tori.typ != get_irn_symconst_attr(b).tori.typ);
1258 static int node_cmp_attr_Call(ir_node *a, ir_node *b)
1260 return (get_irn_call_attr(a) != get_irn_call_attr(b));
1263 static int node_cmp_attr_FuncCall(ir_node *a, ir_node *b)
1265 return (get_irn_funccall_attr(a) != get_irn_funccall_attr(b));
1268 static int node_cmp_attr_Sel(ir_node *a, ir_node *b)
1270 return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind)
1271 || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name)
1272 || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner)
1273 || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name)
1274 || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type);
1277 static int node_cmp_attr_Phi(ir_node *a, ir_node *b)
1279 return get_irn_phi_attr (a) != get_irn_phi_attr (b);
1282 static int node_cmp_attr_Cast(ir_node *a, ir_node *b)
1284 return get_Cast_type(a) != get_Cast_type(b);
1288 * set the default node attribute compare operation
1290 static ir_op *firm_set_default_node_cmp_attr(ir_op *op)
1294 op->node_cmp_attr = node_cmp_attr_##a; \
1308 op->node_cmp_attr = NULL;
1316 * Compare function for two nodes in the hash table. Gets two
1317 * nodes as parameters. Returns 0 if the nodes are a cse.
1320 vt_cmp (const void *elt, const void *key)
1328 if (a == b) return 0;
1330 if ((get_irn_op(a) != get_irn_op(b)) ||
1331 (get_irn_mode(a) != get_irn_mode(b))) return 1;
1333 /* compare if a's in and b's in are equal */
1334 irn_arity_a = get_irn_arity (a);
1335 if (irn_arity_a != get_irn_arity(b))
1338 /* for block-local cse and pinned nodes: */
1339 if (!get_opt_global_cse() || (get_op_pinned(get_irn_op(a)) == pinned)) {
1340 if (get_irn_n(a, -1) != get_irn_n(b, -1))
1344 /* compare a->in[0..ins] with b->in[0..ins] */
1345 for (i = 0; i < irn_arity_a; i++)
1346 if (get_irn_n(a, i) != get_irn_n(b, i))
1350 * here, we already now that the nodes are identical except their
1353 if (a->op->node_cmp_attr)
1354 return a->op->node_cmp_attr(a, b);
1360 * Calculate a hash value of a node.
1363 ir_node_hash (ir_node *node)
1368 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
1369 h = irn_arity = get_irn_arity(node);
1371 /* consider all in nodes... except the block. */
1372 for (i = 0; i < irn_arity; i++) {
1373 h = 9*h + (unsigned long)get_irn_n(node, i);
1377 h = 9*h + (unsigned long) get_irn_mode (node);
1379 h = 9*h + (unsigned long) get_irn_op (node);
1385 new_identities (void)
1387 return new_pset (vt_cmp, N_IR_NODES);
1391 del_identities (pset *value_table)
1393 del_pset (value_table);
1397 * Return the canonical node computing the same value as n.
1398 * Looks up the node in a hash table.
1400 static INLINE ir_node *
1401 identify (pset *value_table, ir_node *n)
1405 if (!value_table) return n;
1407 /* TODO: use a generic commutative attribute */
1408 if (get_opt_reassociation()) {
1409 if (is_op_commutative(get_irn_op(n))) {
1410 /* for commutative operators perform a OP b == b OP a */
1411 if (get_binop_left(n) > get_binop_right(n)) {
1412 ir_node *h = get_binop_left(n);
1413 set_binop_left(n, get_binop_right(n));
1414 set_binop_right(n, h);
1419 o = pset_find (value_table, n, ir_node_hash (n));
1426 * During construction we set the pinned flag in the graph right when the
1427 * optimizatin is performed. The flag turning on procedure global cse could
1428 * be changed between two allocations. This way we are safe.
1430 static INLINE ir_node *
1431 identify_cons (pset *value_table, ir_node *n) {
1433 n = identify(value_table, n);
1434 if (get_irn_n(old, -1) != get_irn_n(n, -1))
1435 set_irg_pinned(current_ir_graph, floats);
1440 * Return the canonical node computing the same value as n.
1441 * Looks up the node in a hash table, enters it in the table
1442 * if it isn't there yet.
1445 identify_remember (pset *value_table, ir_node *node)
1449 if (!value_table) return node;
1451 /* lookup or insert in hash table with given hash key. */
1452 o = pset_insert (value_table, node, ir_node_hash (node));
1454 if (o == node) return node;
1460 add_identities (pset *value_table, ir_node *node) {
1461 identify_remember (value_table, node);
1465 * garbage in, garbage out. If a node has a dead input, i.e., the
1466 * Bad node is input to the node, return the Bad node.
1468 static INLINE ir_node *
1469 gigo (ir_node *node)
1472 ir_op* op = get_irn_op(node);
1474 /* remove garbage blocks by looking at control flow that leaves the block
1475 and replacing the control flow by Bad. */
1476 if (get_irn_mode(node) == mode_X) {
1477 ir_node *block = get_nodes_block(node);
1478 if (op == op_End) return node; /* Don't optimize End, may have Bads. */
1479 if (get_irn_op(block) == op_Block && get_Block_matured(block)) {
1480 irn_arity = get_irn_arity(block);
1481 for (i = 0; i < irn_arity; i++) {
1482 if (!is_Bad(get_irn_n(block, i))) break;
1484 if (i == irn_arity) return new_Bad();
1488 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
1489 blocks predecessors is dead. */
1490 if ( op != op_Block && op != op_Phi && op != op_Tuple) {
1491 irn_arity = get_irn_arity(node);
1492 for (i = -1; i < irn_arity; i++) {
1493 if (is_Bad(get_irn_n(node, i))) {
1499 /* With this code we violate the agreement that local_optimize
1500 only leaves Bads in Block, Phi and Tuple nodes. */
1501 /* If Block has only Bads as predecessors it's garbage. */
1502 /* If Phi has only Bads as predecessors it's garbage. */
1503 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
1504 irn_arity = get_irn_arity(node);
1505 for (i = 0; i < irn_arity; i++) {
1506 if (!is_Bad(get_irn_n(node, i))) break;
1508 if (i == irn_arity) node = new_Bad();
1516 * These optimizations deallocate nodes from the obstack.
1517 * It can only be called if it is guaranteed that no other nodes
1518 * reference this one, i.e., right after construction of a node.
1521 optimize_node (ir_node *n)
1525 opcode iro = get_irn_opcode(n);
1527 /* Allways optimize Phi nodes: part of the construction. */
1528 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
1530 /* constant expression evaluation / constant folding */
1531 if (get_opt_constant_folding()) {
1532 /* constants can not be evaluated */
1533 if (get_irn_op(n) != op_Const) {
1534 /* try to evaluate */
1535 tv = computed_value (n);
1536 if ((get_irn_mode(n) != mode_T) && (tv != tarval_bad)) {
1537 /* evaluation was succesful -- replace the node. */
1538 obstack_free (current_ir_graph->obst, n);
1539 return new_Const (get_tarval_mode (tv), tv);
1544 /* remove unnecessary nodes */
1545 if (get_opt_constant_folding() ||
1546 (iro == iro_Phi) || /* always optimize these nodes. */
1548 (iro == iro_Proj) ||
1549 (iro == iro_Block) ) /* Flags tested local. */
1550 n = equivalent_node (n);
1552 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
1554 /** common subexpression elimination **/
1555 /* Checks whether n is already available. */
1556 /* The block input is used to distinguish different subexpressions. Right
1557 now all nodes are pinned to blocks, i.e., the cse only finds common
1558 subexpressions within a block. */
1560 n = identify_cons (current_ir_graph->value_table, n);
1563 /* We found an existing, better node, so we can deallocate the old node. */
1564 obstack_free (current_ir_graph->obst, old_n);
1567 /* Some more constant expression evaluation that does not allow to
1569 iro = get_irn_opcode(n);
1570 if (get_opt_constant_folding() ||
1571 (iro == iro_Cond) ||
1572 (iro == iro_Proj)) /* Flags tested local. */
1573 n = transform_node (n);
1575 /* Remove nodes with dead (Bad) input.
1576 Run always for transformation induced Bads. */
1579 /* Now we have a legal, useful node. Enter it in hash table for cse */
1580 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
1581 n = identify_remember (current_ir_graph->value_table, n);
1589 * These optimizations never deallocate nodes. This can cause dead
1590 * nodes lying on the obstack. Remove these by a dead node elimination,
1591 * i.e., a copying garbage collection.
1594 optimize_in_place_2 (ir_node *n)
1598 opcode iro = get_irn_opcode(n);
1600 if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n;
1602 /* if not optimize return n */
1605 /* Here this is possible. Why? */
1610 /* constant expression evaluation / constant folding */
1611 if (get_opt_constant_folding()) {
1612 /* constants can not be evaluated */
1613 if (iro != iro_Const) {
1614 /* try to evaluate */
1615 tv = computed_value (n);
1616 if ((get_irn_mode(n) != mode_T) && (tv != tarval_bad)) {
1617 /* evaluation was succesful -- replace the node. */
1618 n = new_Const (get_tarval_mode (tv), tv);
1619 __dbg_info_merge_pair(n, old_n, dbg_const_eval);
1625 /* remove unnecessary nodes */
1626 /*if (get_opt_constant_folding()) */
1627 if (get_opt_constant_folding() ||
1628 (iro == iro_Phi) || /* always optimize these nodes. */
1629 (iro == iro_Id) || /* ... */
1630 (iro == iro_Proj) || /* ... */
1631 (iro == iro_Block) ) /* Flags tested local. */
1632 n = equivalent_node (n);
1634 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
1636 /** common subexpression elimination **/
1637 /* Checks whether n is already available. */
1638 /* The block input is used to distinguish different subexpressions. Right
1639 now all nodes are pinned to blocks, i.e., the cse only finds common
1640 subexpressions within a block. */
1641 if (get_opt_cse()) {
1642 n = identify (current_ir_graph->value_table, n);
1645 /* Some more constant expression evaluation. */
1646 iro = get_irn_opcode(n);
1647 if (get_opt_constant_folding() ||
1648 (iro == iro_Cond) ||
1649 (iro == iro_Proj)) /* Flags tested local. */
1650 n = transform_node (n);
1652 /* Remove nodes with dead (Bad) input.
1653 Run always for transformation induced Bads. */
1656 /* Now we can verify the node, as it has no dead inputs any more. */
1659 /* Now we have a legal, useful node. Enter it in hash table for cse.
1660 Blocks should be unique anyways. (Except the successor of start:
1661 is cse with the start block!) */
1662 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
1663 n = identify_remember (current_ir_graph->value_table, n);
1669 * Wrapper for external use, set proper status bits after optimization.
1672 optimize_in_place (ir_node *n)
1674 /* Handle graph state */
1675 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1677 if (get_opt_global_cse())
1678 set_irg_pinned(current_ir_graph, floats);
1679 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
1680 set_irg_outs_inconsistent(current_ir_graph);
1681 /* Maybe we could also test whether optimizing the node can
1682 change the control graph. */
1683 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
1684 set_irg_dom_inconsistent(current_ir_graph);
1685 return optimize_in_place_2 (n);
1689 * set the default ir op operations
1691 ir_op *firm_set_default_operations(ir_op *op)
1693 op = firm_set_default_computed_value(op);
1694 op = firm_set_default_equivalent_node(op);
1695 op = firm_set_default_transform_node(op);
1696 op = firm_set_default_node_cmp_attr(op);