1 /* Copyright (C) 1998 - 2000 by Universitaet Karlsruhe
2 ** All rights reserved.
4 ** Authors: Christian Schaefer, Goetz Lindenmaier
6 ** iropt --- optimizations intertwined with IR construction.
15 # include "irnode_t.h"
16 # include "irgraph_t.h"
25 /* Make types visible to allow most efficient access */
26 # include "entity_t.h"
28 /* Trivial inlineable routine for copy propagation.
29 Does follow Ids, needed to optimize inlined code. */
30 static inline ir_node *
31 follow_Id (ir_node *n)
33 while (get_irn_op (n) == op_Id) n = get_Id_pred (n);
37 static inline tarval *
40 if ((n != NULL) && (get_irn_op(n) == op_Const))
41 return get_Const_tarval(n);
46 /* if n can be computed, return the value, else NULL. Performs
47 constant folding. GL: Only if n is arithmetic operator? */
49 computed_value (ir_node *n)
53 ir_node *a = NULL, *b = NULL; /* initialized to shut up gcc */
54 tarval *ta = NULL, *tb = NULL; /* initialized to shut up gcc */
58 /* get the operands we will work on for simple cases. */
60 a = get_binop_left(n);
61 b = get_binop_right(n);
62 } else if (is_unop(n)) {
66 /* if the operands are constants, get the target value, else set it NULL.
67 (a and b may be NULL if we treat a node that is no computation.) */
71 /* Perform the constant evaluation / computation. */
72 switch (get_irn_opcode(n)) {
74 res = get_Const_tarval(n);
76 if (ta && tb && (get_irn_mode(a) == get_irn_mode(b))
77 && (get_irn_mode(a) != mode_p)) {
78 res = tarval_add (ta, tb);
82 if (ta && tb && (get_irn_mode(a) == get_irn_mode(b))
83 && (get_irn_mode(a) != mode_p)) {
84 res = tarval_sub (ta, tb);
86 res = tarval_mode_null [get_irn_modecode (n)];
90 if (ta && mode_is_float(get_irn_mode(a)))
91 res = tarval_neg (ta);
94 if (ta && tb) /* tarval_mul tests for equivalent modes itself */ {
95 res = tarval_mul (ta, tb);
97 /* a*0 = 0 or 0*b = 0:
98 calls computed_value recursive and returns the 0 with proper
101 if ( (tarval_classify ((v = computed_value (a))) == 0)
102 || (tarval_classify ((v = computed_value (b))) == 0)) {
108 /* This was missing in original implementation. Why? */
109 if (ta && tb && (get_irn_mode(a) == get_irn_mode(b))) {
110 if (tarval_classify(tb) == 0) {res = NULL; break;}
111 res = tarval_quo(ta, tb);
115 /* This was missing in original implementation. Why? */
116 if (ta && tb && (get_irn_mode(a) == get_irn_mode(b))) {
117 if (tarval_classify(tb) == 0) {res = NULL; break;}
118 res = tarval_div(ta, tb);
122 /* This was missing in original implementation. Why? */
123 if (ta && tb && (get_irn_mode(a) == get_irn_mode(b))) {
124 if (tarval_classify(tb) == 0) {res = NULL; break;}
125 res = tarval_mod(ta, tb);
128 /* for iro_DivMod see iro_Proj */
131 res = tarval_abs (ta);
135 res = tarval_and (ta, tb);
138 if ( (tarval_classify ((v = computed_value (a))) == 0)
139 || (tarval_classify ((v = computed_value (b))) == 0)) {
146 res = tarval_or (ta, tb);
149 if ( (tarval_classify ((v = computed_value (a))) == -1)
150 || (tarval_classify ((v = computed_value (b))) == -1)) {
155 case iro_Eor: if (ta && tb) { res = tarval_eor (ta, tb); } break;
156 case iro_Not: if (ta) { res = tarval_neg (ta); } break;
157 case iro_Shl: if (ta && tb) { res = tarval_shl (ta, tb); } break;
158 /* tarval_shr is faulty !! */
159 case iro_Shr: if (ta && tb) { res = tarval_shr (ta, tb); } break;
160 case iro_Shrs:if (ta && tb) { /*res = tarval_shrs (ta, tb)*/; } break;
161 case iro_Rot: if (ta && tb) { /*res = tarval_rot (ta, tb)*/; } break;
162 case iro_Conv:if (ta) { res = tarval_convert_to (ta, get_irn_mode (n)); }
164 case iro_Proj: /* iro_Cmp */
168 a = get_Proj_pred(n);
169 /* Optimize Cmp nodes.
170 This performs a first step of unreachable code elimination.
171 Proj can not be computed, but folding a Cmp above the Proj here is
172 not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
174 There are several case where we can evaluate a Cmp node:
175 1. The nodes compared are both the same. If we compare for
176 equal, this will return true, else it will return false.
177 This step relies on cse.
178 2. The predecessors of Cmp are target values. We can evaluate
180 3. The predecessors are Allocs or void* constants. Allocs never
181 return NULL, they raise an exception. Therefore we can predict
183 if (get_irn_op(a) == op_Cmp) {
184 aa = get_Cmp_left(a);
185 ab = get_Cmp_right(a);
186 if (aa == ab) { /* 1.: */
187 /* This is a tric with the bits used for encoding the Cmp
188 Proj numbers, the following statement is not the same:
189 res = tarval_from_long (mode_b, (get_Proj_proj(n) == Eq)): */
190 res = tarval_from_long (mode_b, (get_Proj_proj(n) & irpn_Eq));
192 tarval *taa = computed_value (aa);
193 tarval *tab = computed_value (ab);
194 if (taa && tab) { /* 2.: */
195 /* strange checks... */
196 ir_pncmp flags = tarval_comp (taa, tab);
197 if (flags != irpn_False) {
198 res = tarval_from_long (mode_b, get_Proj_proj(n) & flags);
200 } else { /* check for 3.: */
201 ir_node *aaa = skip_nop(skip_Proj(aa));
202 ir_node *aba = skip_nop(skip_Proj(ab));
203 if ( ( (/* aa is ProjP and aaa is Alloc */
204 (get_irn_op(aa) == op_Proj)
205 && (get_irn_mode(aa) == mode_p)
206 && (get_irn_op(aaa) == op_Alloc))
207 && ( (/* ab is constant void */
208 (get_irn_op(ab) == op_Const)
209 && (get_irn_mode(ab) == mode_p)
210 && (get_Const_tarval(ab) == tarval_p_void))
211 || (/* ab is other Alloc */
212 (get_irn_op(ab) == op_Proj)
213 && (get_irn_mode(ab) == mode_p)
214 && (get_irn_op(aba) == op_Alloc)
216 || (/* aa is void and aba is Alloc */
217 (get_irn_op(aa) == op_Const)
218 && (get_irn_mode(aa) == mode_p)
219 && (get_Const_tarval(aa) == tarval_p_void)
220 && (get_irn_op(ab) == op_Proj)
221 && (get_irn_mode(ab) == mode_p)
222 && (get_irn_op(aba) == op_Alloc)))
224 res = tarval_from_long (mode_b, get_Proj_proj(n) & irpn_Ne);
227 } else if (get_irn_op(a) == op_DivMod) {
228 ta = value_of(get_DivMod_left(a));
229 tb = value_of(get_DivMod_right(a));
230 if (ta && tb && (get_irn_mode(a) == get_irn_mode(b))) {
231 if (tarval_classify(tb) == 0) {res = NULL; break;}
232 if (get_Proj_proj(n)== 0) /* Div */
233 res = tarval_div(ta, tb);
235 res = tarval_mod(ta, tb);
238 /* printf(" # comp_val: Proj node, not optimized\n"); */
250 /* returns 1 if the a and b are pointers to different locations. */
252 different_identity (ir_node *a, ir_node *b)
254 assert (get_irn_mode (a) == mode_p
255 && get_irn_mode (b) == mode_p);
257 if (get_irn_op (a) == op_Proj && get_irn_op(b) == op_Proj) {
258 ir_node *a1 = get_Proj_pred (a);
259 ir_node *b1 = get_Proj_pred (b);
260 if (a1 != b1 && get_irn_op (a1) == op_Alloc
261 && get_irn_op (b1) == op_Alloc)
268 /* equivalent_node returns a node equivalent to N. It skips all nodes that
269 perform no actual computation, as, e.g., the Id nodes. It does not create
270 new nodes. It is therefore safe to free N if the node returned is not N.
271 If a node returns a Tuple we can not just skip it. If the size of the
272 in array fits, we transform n into a tuple (e.g., Div). */
274 equivalent_node (ir_node *n)
277 ir_node *a = NULL; /* to shutup gcc */
278 ir_node *b = NULL; /* to shutup gcc */
279 ir_node *c = NULL; /* to shutup gcc */
281 ins = get_irn_arity (n);
283 /* get the operands we will work on */
285 a = get_binop_left(n);
286 b = get_binop_right(n);
287 } else if (is_unop(n)) {
291 /* skip unnecessary nodes. */
292 switch (get_irn_opcode (n)) {
295 /* The Block constructor does not call optimize, but mature_block
296 calls the optimization. */
297 assert(get_Block_matured(n));
299 /* A single entry Block following a single exit Block can be merged,
300 if it is not the Start block. */
301 /* !!! Beware, all Phi-nodes of n must have been optimized away.
302 This should be true, as the block is matured before optimize is called.
303 But what about Phi-cycles with the Phi0/Id that could not be resolved?
304 Remaining Phi nodes are just Ids. */
305 if (get_Block_n_cfgpreds(n) == 1
306 && get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp) {
307 n = get_nodes_Block(get_Block_cfgpred(n, 0));
309 } else if ((n != current_ir_graph->start_block) &&
310 (n != current_ir_graph->end_block) ) {
312 /* If all inputs are dead, this block is dead too, except if it is
313 the start or end block. This is a step of unreachable code
315 for (i = 0; i < get_Block_n_cfgpreds(n); i++) {
316 if (!is_Bad(get_Block_cfgpred(n, i))) break;
318 if (i == get_Block_n_cfgpreds(n))
324 case iro_Jmp: /* GL: Why not same for op_Raise?? */
325 /* unreachable code elimination */
326 if (is_Bad(get_nodes_Block(n))) n = new_Bad();
328 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
329 See cases for iro_Cond and iro_Proj in transform_node. */
330 /** remove stuff as x+0, x*1 x&true ... constant expression evaluation **/
331 case iro_Or: if (a == b) {n = a; break;}
336 /* After running compute_node there is only one constant predecessor.
337 Find this predecessors value and remember the other node: */
338 if ((tv = computed_value (a))) {
340 } else if ((tv = computed_value (b))) {
344 /* If this predecessors constant value is zero, the operation is
345 unnecessary. Remove it: */
346 if (tarval_classify (tv) == 0) {
356 /* these operations are not commutative. Test only one predecessor. */
357 if (tarval_classify (computed_value (b)) == 0) {
359 /* Test if b > #bits of a ==> return 0 / divide b by #bits
360 --> transform node? */
363 case iro_Not: /* NotNot x == x */
364 case iro_Minus: /* --x == x */ /* ??? Is this possible or can --x raise an
365 out of bounds exception if min =! max? */
366 if (get_irn_op(get_unop_op(n)) == get_irn_op(n))
367 n = get_unop_op(get_unop_op(n));
370 /* Mul is commutative and has again an other neutral element. */
371 if (tarval_classify (computed_value (a)) == 1) {
373 } else if (tarval_classify (computed_value (b)) == 1) {
378 /* Div is not commutative. */
379 if (tarval_classify (computed_value (b)) == 1) { /* div(x, 1) == x */
380 /* Turn Div into a tuple (mem, bad, a) */
381 ir_node *mem = get_Div_mem(n);
382 turn_into_tuple(n, 3);
383 set_Tuple_pred(n, 0, mem);
384 set_Tuple_pred(n, 1, new_Bad());
385 set_Tuple_pred(n, 2, a);
388 /* GL: Why are they skipped? DivMod allocates new nodes --> it's
389 teated in transform node.
390 case iro_Mod, Quot, DivMod
394 /* And has it's own neutral element */
395 else if (tarval_classify (computed_value (a)) == -1) {
397 } else if (tarval_classify (computed_value (b)) == -1) {
402 if (get_irn_mode(n) == get_irn_mode(a)) { /* No Conv necessary */
404 } else if (get_irn_mode(n) == mode_b) {
405 if (get_irn_op(a) == op_Conv &&
406 get_irn_mode (get_Conv_op(a)) == mode_b) {
407 n = get_Conv_op(a); /* Convb(Conv*(xxxb(...))) == xxxb(...) */
414 /* Several optimizations:
415 - no Phi in start block.
416 - remove Id operators that are inputs to Phi
417 - fold Phi-nodes, iff they have only one predecessor except
423 ir_node *block = NULL; /* to shutup gcc */
424 ir_node *first_val = NULL; /* to shutup gcc */
425 ir_node *scnd_val = NULL; /* to shutup gcc */
427 n_preds = get_Phi_n_preds(n);
429 block = get_nodes_Block(n);
430 if (is_Bad(block)) return new_Bad();
431 assert(get_irn_op (block) == op_Block);
433 /* there should be no Phi nodes in the Start region. */
434 if (block == current_ir_graph->start_block) {
439 if (n_preds == 0) { /* Phi of dead Region without predecessors. */
440 /* GL: why not return new_Bad? */
445 /* first we test for a special case: */
446 /* Confirm is a special node fixing additional information for a
447 value that is known at a certain point. This is useful for
448 dataflow analysis. */
450 ir_node *a = follow_Id (get_Phi_pred(n, 0));
451 ir_node *b = follow_Id (get_Phi_pred(n, 1));
452 if ( (get_irn_op(a) == op_Confirm)
453 && (get_irn_op(b) == op_Confirm)
454 && (follow_Id (get_irn_n(a, 0)) == follow_Id(get_irn_n(b, 0)))
455 && (get_irn_n(a, 1) == get_irn_n (b, 1))
456 && (a->data.num == (~b->data.num & irpn_True) )) {
457 n = follow_Id (get_irn_n(a, 0));
463 /* Find first non-self-referencing input */
464 for (i = 0; i < n_preds; ++i) {
465 first_val = follow_Id(get_Phi_pred(n, i));
467 set_Phi_pred(n, i, first_val);
468 if ( (first_val != n) /* not self pointer */
469 && (get_irn_op(first_val) != op_Bad) /* value not dead */
470 && !(is_Bad (get_Block_cfgpred(block, i))) ) { /* not dead control flow */
471 break; /* then found first value. */
475 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
476 if (i >= n_preds) { n = new_Bad(); break; }
480 /* follow_Id () for rest of inputs, determine if any of these
481 are non-self-referencing */
482 while (++i < n_preds) {
483 scnd_val = follow_Id(get_Phi_pred(n, i));
485 set_Phi_pred(n, i, scnd_val);
487 && (scnd_val != first_val)
488 && (get_irn_op(scnd_val) != op_Bad)
489 && !(is_Bad (get_Block_cfgpred(block, i))) ) {
494 /* Fold, if no multiple distinct non-self-referencing inputs */
498 /* skip the remaining Ids. */
499 while (++i < n_preds) {
500 set_Phi_pred(n, i, follow_Id(get_Phi_pred(n, i)));
508 #if 0 /* Is an illegal transformation: different nodes can
509 represent the same pointer value!! */
510 a = skip_Proj(get_Load_mem(n));
513 if (get_irn_op(a) == op_Store) {
514 if ( different_identity (b, get_Store_ptr(a))) {
515 /* load and store use different pointers, therefore load
516 needs not take store's memory but the state before. */
517 set_Load_mem (n, get_Store_mem(a));
518 } else if (( 0 /* ???didn't get cryptic test that returns 0 */ )) {
525 /* remove unnecessary store. */
527 a = skip_Proj(get_Store_mem(n));
528 b = get_Store_ptr(n);
529 c = skip_Proj(get_Store_value(n));
531 if (get_irn_op(a) == op_Store
532 && get_Store_ptr(a) == b
533 && skip_Proj(get_Store_value(a)) == c) {
534 /* We have twice exactly the same store -- a write after write. */
536 } else if (get_irn_op(c) == op_Load
537 && (a == c || skip_Proj(get_Load_mem(c)) == a)
538 && get_Load_ptr(c) == b )
539 /* !!!??? and a cryptic test */ {
540 /* We just loaded the value from the same memory, i.e., the store
541 doesn't change the memory -- a write after read. */
542 a = get_Store_mem(n);
543 turn_into_tuple(n, 2);
544 set_Tuple_pred(n, 0, a);
545 set_Tuple_pred(n, 1, new_Bad());
552 a = get_Proj_pred(n);
554 if ( get_irn_op(a) == op_Tuple) {
555 /* Remove the Tuple/Proj combination. */
556 if ( get_Proj_proj(n) <= get_Tuple_n_preds(a) ) {
557 n = get_Tuple_pred(a, get_Proj_proj(n));
559 assert(0); /* This should not happen! */
562 } else if (get_irn_mode(n) == mode_X &&
563 is_Bad(get_nodes_Block(n))) {
564 /* Remove dead control flow. */
578 } /* end equivalent_node() */
581 /* tries several [inplace] [optimizing] transformations and returns a
582 equivalent node. The difference to equivalent_node is that these
583 transformations _do_ generate new nodes, and thus the old node must
584 not be freed even if the equivalent node isn't the old one. */
586 transform_node (ir_node *n)
589 ir_node *a = NULL, *b;
592 switch (get_irn_opcode(n)) {
598 a = get_DivMod_left(n);
599 b = get_DivMod_right(n);
600 mode = get_irn_mode(a);
602 if (!( mode_is_int(get_irn_mode(a))
603 && mode_is_int(get_irn_mode(b))))
607 a = new_Const (mode, tarval_from_long (mode, 1));
608 b = new_Const (mode, tarval_from_long (mode, 0));
615 if (tarval_classify(tb) == 1) {
616 b = new_Const (mode, tarval_from_long (mode, 0));
620 resa = tarval_div (ta, tb);
621 if (!resa) break; /* Causes exception!!! Model by replacing through
622 Jmp for X result!? */
623 resb = tarval_mod (ta, tb);
624 if (!resb) break; /* Causes exception! */
625 a = new_Const (mode, resa);
626 b = new_Const (mode, resb);
629 } else if (tarval_classify (ta) == 0) {
634 if (evaluated) { /* replace by tuple */
635 ir_node *mem = get_DivMod_mem(n);
636 turn_into_tuple(n, 4);
637 set_Tuple_pred(n, 0, mem);
638 set_Tuple_pred(n, 1, new_Bad()); /* no exception */
639 set_Tuple_pred(n, 2, a);
640 set_Tuple_pred(n, 3, b);
641 assert(get_nodes_Block(n));
647 /* Replace the Cond by a Jmp if it branches on a constant
650 a = get_Cond_selector(n);
653 if (ta && (get_irn_mode(a) == mode_b)) {
654 /* It's a boolean Cond, branching on a boolean constant.
655 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
656 jmp = new_r_Jmp(current_ir_graph, get_nodes_Block(n));
657 turn_into_tuple(n, 2);
658 if (tv_val_b(ta) == 1) /* GL: I hope this returns 1 if true */ {
659 set_Tuple_pred(n, 0, new_Bad());
660 set_Tuple_pred(n, 1, jmp);
662 set_Tuple_pred(n, 0, jmp);
663 set_Tuple_pred(n, 1, new_Bad());
665 /* We might generate an endless loop, so keep it alive. */
666 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_Block(n));
667 } else if (ta && (get_irn_mode(a) == mode_I) && (get_Cond_kind(n) == dense)) {
668 /* I don't want to allow Tuples smaller than the biggest Proj.
669 Also this tuple might get really big...
670 I generate the Jmp here, and remember it in link. Link is used
671 when optimizing Proj. */
672 set_irn_link(n, new_r_Jmp(current_ir_graph, get_nodes_Block(n)));
673 /* We might generate an endless loop, so keep it alive. */
674 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_Block(n));
675 } else if ( (get_irn_op(get_Cond_selector(n)) == op_Eor)
676 && (get_irn_mode(get_Cond_selector(n)) == mode_b)
677 && (tarval_classify(computed_value(get_Eor_right(a))) == 1)) {
678 /* The Eor is a negate. Generate a new Cond without the negate,
679 simulate the negate by exchanging the results. */
680 set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_Block(n),
682 } else if ( (get_irn_op(get_Cond_selector(n)) == op_Not)
683 && (get_irn_mode(get_Cond_selector(n)) == mode_b)) {
684 /* A Not before the Cond. Generate a new Cond without the Not,
685 simulate the Not by exchanging the results. */
686 set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_Block(n),
693 a = get_Proj_pred(n);
695 if ( (get_irn_op(a) == op_Cond)
697 && get_irn_op(get_irn_link(a)) == op_Cond) {
698 /* Use the better Cond if the Proj projs from a Cond which get's
699 its result from an Eor/Not. */
700 assert ( ( (get_irn_op(get_Cond_selector(a)) == op_Eor)
701 || (get_irn_op(get_Cond_selector(a)) == op_Not))
702 && (get_irn_mode(get_Cond_selector(a)) == mode_b)
703 && (get_irn_op(get_irn_link(a)) == op_Cond)
704 && (get_Cond_selector(get_irn_link(a)) ==
705 get_Eor_left(get_Cond_selector(a))));
706 set_Proj_pred(n, get_irn_link(a));
707 if (get_Proj_proj(n) == 0)
711 } else if ( (get_irn_op(a) == op_Cond)
712 && (get_irn_mode(get_Cond_selector(a)) == mode_I)
714 && (get_Cond_kind(a) == dense)) {
715 /* The Cond is a Switch on a Constant */
716 if (get_Proj_proj(n) == tv_val_CHIL(value_of(a))) {
717 /* The always taken branch, reuse the existing Jmp. */
718 if (!get_irn_link(a)) /* well, if it exists ;-> */
719 set_irn_link(a, new_r_Jmp(current_ir_graph, get_nodes_Block(n)));
720 assert(get_irn_op(get_irn_link(a)) == op_Jmp);
722 } else {/* Not taken control flow, but be careful with the default! */
723 if (get_Proj_proj(n) < a->attr.c.default_proj){
724 /* a never taken branch */
727 a->attr.c.default_proj = get_Proj_proj(n);
732 case iro_Eor: { /* @@@ not tested as boolean Eor not allowed any more. */
734 b = get_Eor_right(n);
736 if ( (get_irn_mode(n) == mode_b)
737 && (get_irn_op(a) == op_Proj)
738 && (get_irn_mode(a) == mode_b)
739 && (tarval_classify (computed_value (b)) == 1)
740 && (get_irn_op(get_Proj_pred(a)) == op_Cmp))
741 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
742 n = new_r_Proj(current_ir_graph, get_nodes_Block(n), get_Proj_pred(a),
743 mode_b, get_negated_pnc(get_Proj_proj(a)));
744 else if ( (get_irn_mode(n) == mode_b)
745 && (tarval_classify (computed_value (b)) == 1))
746 /* The Eor is a Not. Replace it by a Not. */
747 /* ????!!!Extend to bitfield 1111111. */
748 n = new_r_Not(current_ir_graph, get_nodes_Block(n), a, mode_b);
754 if ( (get_irn_mode(n) == mode_b)
755 && (get_irn_op(a) == op_Proj)
756 && (get_irn_mode(a) == mode_b)
757 && (get_irn_op(get_Proj_pred(a)) == op_Cmp))
758 /* We negate a Cmp. The Cmp has the negated result anyways! */
759 n = new_r_Proj(current_ir_graph, get_nodes_Block(n), get_Proj_pred(a),
760 mode_b, get_negated_pnc(get_Proj_proj(a)));
768 /* **************** Common Subexpression Elimination **************** */
770 /* Compare function for two nodes in the hash table. Gets two */
771 /* nodes as parameters. Returns 0 if the nodes are a cse. */
773 vt_cmp (const void *elt, const void *key)
781 if (a == b) return 0;
783 if ((get_irn_op(a) != get_irn_op(b)) ||
784 (get_irn_mode(a) != get_irn_mode(b))) return 1;
786 /* compare if a's in and b's in are equal */
787 /* GL: we optimize only nodes with in arrays of fixed sizes.
788 if (get_irn_arity (a) != -2) {
789 ins = get_irn_arity (a);
790 if (ins != get_irn_arity (b)) return 1;
791 ain = get_irn_in (a);
792 bin = get_irn_in (b);
795 if (get_irn_arity (a) != get_irn_arity(b))
798 /* for block-local cse and pinned nodes: */
799 if (!get_opt_global_cse() || (get_op_pinned(get_irn_op(a)) == pinned)) {
800 if (get_irn_n(a, -1) != get_irn_n(b, -1))
804 /* compare a->in[0..ins] with b->in[0..ins] */
805 for (i = 0; i < get_irn_arity(a); i++)
806 if (get_irn_n(a, i) != get_irn_n(b, i))
809 switch (get_irn_opcode(a)) {
811 return get_irn_const_attr (a) != get_irn_const_attr (b);
813 return get_irn_proj_attr (a) != get_irn_proj_attr (b);
815 return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where)
816 || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type);
818 return (get_irn_free_attr(a) != get_irn_free_attr(b));
820 return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num)
821 || (get_irn_symconst_attr(a).tori.typ != get_irn_symconst_attr(b).tori.typ);
823 return (get_irn_call_attr(a) != get_irn_call_attr(b));
825 return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind)
826 || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name)
827 || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner)
828 || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name)
829 || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type)
830 || (get_irn_sel_attr(a).ltyp != get_irn_sel_attr(b).ltyp);
832 return get_irn_phi_attr (a) != get_irn_phi_attr (b);
840 ir_node_hash (ir_node *node)
845 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
846 h = get_irn_arity(node);
848 /* consider all in nodes... except the block. */
849 for (i = 0; i < get_irn_arity(node); i++) {
850 h = 9*h + (unsigned long)get_irn_n(node, i);
854 h = 9*h + (unsigned long) get_irn_mode (node);
856 h = 9*h + (unsigned long) get_irn_op (node);
862 new_identities (void)
864 return new_pset (vt_cmp, TUNE_NIR_NODES);
868 del_identities (pset *value_table)
870 del_pset (value_table);
873 /* Return the canonical node computing the same value as n.
874 Looks up the node in a hash table. */
875 static inline ir_node *
876 identify (pset *value_table, ir_node *n)
880 if (!value_table) return n;
882 if (get_opt_reassociation()) {
883 switch (get_irn_opcode (n)) {
890 /* for commutative operators perform a OP b == b OP a */
891 if (get_binop_left(n) > get_binop_right(n)) {
892 ir_node *h = get_binop_left(n);
893 set_binop_left(n, get_binop_right(n));
894 set_binop_right(n, h);
902 o = pset_find (value_table, n, ir_node_hash (n));
908 /* During construction we set the pinned flag in the graph right when the
909 optimizatin is performed. The flag turning on procedure global cse could
910 be changed between two allocations. This way we are safe. */
911 static inline ir_node *
912 identify_cons (pset *value_table, ir_node *n) {
914 n = identify(value_table, n);
915 if (get_irn_n(old, -1) != get_irn_n(n, -1))
916 set_irg_pinned(current_ir_graph, floats);
920 /* Return the canonical node computing the same value as n.
921 Looks up the node in a hash table, enters it in the table
922 if it isn't there yet. */
924 identify_remember (pset *value_table, ir_node *node)
928 if (!value_table) return node;
930 /* lookup or insert in hash table with given hash key. */
931 o = pset_insert (value_table, node, ir_node_hash (node));
933 if (o == node) return node;
939 add_identities (pset *value_table, ir_node *node) {
940 identify_remember (value_table, node);
943 /* garbage in, garbage out. If a node has a dead input, i.e., the
944 Bad node is input to the node, return the Bad node. */
945 static inline ir_node *
949 ir_op* op = get_irn_op(node);
951 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
952 blocks predecessors is dead. */
953 if ( op != op_Block && op != op_Phi && op != op_Tuple) {
954 for (i = -1; i < get_irn_arity(node); i++) {
955 if (is_Bad(get_irn_n(node, i))) {
961 /* If Block has only Bads as predecessors it's garbage. */
962 /* If Phi has only Bads as predecessors it's garbage. */
963 if (op == op_Block || op == op_Phi) {
964 for (i = 0; i < get_irn_arity(node); i++) {
965 if (!is_Bad(get_irn_n(node, i))) break;
967 if (i = get_irn_arity(node)) node = new_Bad();
974 /* These optimizations deallocate nodes from the obstack.
975 It can only be called if it is guaranteed that no other nodes
976 reference this one, i.e., right after construction of a node. */
978 optimize (ir_node *n)
983 /* Allways optimize Phi nodes: part of the construction. */
984 if ((!get_optimize()) && (get_irn_op(n) != op_Phi)) return n;
986 /* constant expression evaluation / constant folding */
987 if (get_opt_constant_folding()) {
988 /* constants can not be evaluated */
989 if (get_irn_op(n) != op_Const) {
990 /* try to evaluate */
991 tv = computed_value (n);
993 /* evaluation was succesful -- replace the node. */
994 obstack_free (current_ir_graph->obst, n);
995 return new_Const (get_tv_mode (tv), tv);
1000 /* remove unnecessary nodes */
1001 if (get_opt_constant_folding() || get_irn_op(n) == op_Phi)
1002 n = equivalent_node (n);
1004 /** common subexpression elimination **/
1005 /* Checks whether n is already available. */
1006 /* The block input is used to distinguish different subexpressions. Right
1007 now all nodes are pinned to blocks, i.e., the cse only finds common
1008 subexpressions within a block. */
1010 n = identify_cons (current_ir_graph->value_table, n);
1011 /* identify found a cse, so deallocate the old node. */
1013 obstack_free (current_ir_graph->obst, old_n);
1014 /* The AmRoq fiasco returns n here. Martin's version doesn't. */
1017 /* Some more constant expression evaluation that does not allow to
1019 if (get_opt_constant_folding())
1020 n = transform_node (n);
1022 /* Remove nodes with dead (Bad) input. */
1023 if (get_opt_unreachable_code())
1025 /* Now we can verify the node, as it has no dead inputs any more. */
1028 /* Now we have a legal, useful node. Enter it in hash table for cse */
1029 if (get_opt_cse()) {
1030 n = identify_remember (current_ir_graph->value_table, n);
1033 #if 0 /* GL: what's the use of this?? */
1034 if ((current_ir_graph->state & irgs_building) && IR_KEEP_ALIVE (n)) {
1035 assert (~current_ir_graph->state & irgs_keep_alives_in_arr);
1036 pdeq_putr (current_ir_graph->keep.living, n);
1043 /* These optimizations never deallocate nodes. This can cause dead
1044 nodes lying on the obstack. Remove these by a dead node elimination,
1045 i.e., a copying garbage collection. */
1047 optimize_in_place_2 (ir_node *n)
1052 if (!get_optimize()) return n;
1054 /* if not optimize return n */
1056 /* Here this is possible. Why? */
1060 /* constant expression evaluation / constant folding */
1061 if (get_opt_constant_folding()) {
1062 /* constants can not be evaluated */
1063 if (get_irn_op(n) != op_Const) {
1064 /* try to evaluate */
1065 tv = computed_value (n);
1067 /* evaluation was succesful -- replace the node. */
1068 n = new_Const (get_tv_mode (tv), tv);
1069 deb_info_copy(n, old_n, id_from_str("const_eval", 10));
1071 /* xprintf("* optimize: computed node %I\n", n->op->name);*/
1076 /* remove unnecessary nodes */
1077 /*if (get_opt_constant_folding()) */
1078 if (get_opt_constant_folding() || get_irn_op(n) == op_Phi)
1079 n = equivalent_node (n);
1081 /** common subexpression elimination **/
1082 /* Checks whether n is already available. */
1083 /* The block input is used to distinguish different subexpressions. Right
1084 now all nodes are pinned to blocks, i.e., the cse only finds common
1085 subexpressions within a block. */
1086 if (get_opt_cse()) {
1087 n = identify (current_ir_graph->value_table, n);
1090 /* identify found a cse, so deallocate the old node. */
1092 /* The AmRoq fiasco returns n here. Martin's version doesn't. */
1095 /* Some more constant expression evaluation. */
1096 if (get_opt_constant_folding())
1097 n = transform_node (n);
1099 /* Remove nodes with dead (Bad) input. */
1100 if (get_opt_unreachable_code())
1102 /* Now we can verify the node, as it has no dead inputs any more. */
1105 /* Now we have a legal, useful node. Enter it in hash table for cse.
1106 Blocks should be unique anyways. (Except the successor of start:
1107 is cse with the start block!) */
1108 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
1109 n = identify_remember (current_ir_graph->value_table, n);
1114 /* Wrapper for external use, set proper status bits after optimization */
1116 optimize_in_place (ir_node *n) {
1117 /* Handle graph state */
1118 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1119 if (get_opt_global_cse())
1120 set_irg_pinned(current_ir_graph, floats);
1121 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
1122 set_irg_outs_inconsistent(current_ir_graph);
1123 return optimize_in_place_2 (n);