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 /* Trivial inlineable routine for copy propagation.
16 Does follow Ids, needed to optimize inlined code. */
17 static inline ir_node *
18 follow_Id (ir_node *n)
20 while (get_irn_op (n) == op_Id) n = get_Id_pred (n);
25 static inline tarval *
28 if ((n != NULL) && (get_irn_op(n) == op_Const))
29 return get_Const_tarval(n);
35 /* if n can be computed, return the value, else NULL. Performs
36 constant Folding. GL: Only if n is arithmetic operator? */
38 computed_value (ir_node *n)
42 ir_node *a = NULL, *b = NULL; /* initialized to shut up gcc */
43 tarval *ta = NULL, *tb = NULL; /* initialized to shut up gcc */
47 /* get the operands we will work on for simple cases. */
49 a = get_binop_left(n);
50 b = get_binop_right(n);
51 } else if (is_unop(n)) {
55 /* if the operands are constants, get the target value, else set it NULL.
56 (a and b may be NULL if we treat a node that is no computation.) */
60 /* Perform the constant evaluation / computation. */
61 switch (get_irn_opcode(n)) {
63 res = get_Const_tarval(n);
65 if (ta && tb && (get_irn_mode(a) == get_irn_mode(b))
66 && (get_irn_mode(a) != mode_p)) {
67 res = tarval_add (ta, tb);
71 if (ta && tb && (get_irn_mode(a) == get_irn_mode(b))
72 && (get_irn_mode(a) != mode_p)) {
73 res = tarval_sub (ta, tb);
75 res = tarval_mode_null [get_irn_modecode (n)];
79 if (ta && mode_is_float(get_irn_mode(a)))
80 res = /*tarval_minus (ta);*/ res;
83 if (ta && tb) /* tarval_mul tests for equivalent modes itself */ {
84 res = tarval_mul (ta, tb);
86 /* calls computed_value recursive and returns the 0 with proper
87 mode. Why is this an extra case? */
89 if ( (tarval_classify ((v = computed_value (a))) == 0)
90 || (tarval_classify ((v = computed_value (b))) == 0)) {
97 res = /*tarval_abs (ta);*/ res;
98 /* allowded or problems with max/min ?? */
102 res = tarval_and (ta, tb);
105 if ( (tarval_classify ((v = computed_value (a))) == 0)
106 || (tarval_classify ((v = computed_value (b))) == 0)) {
113 res = tarval_or (ta, tb);
116 if ( (tarval_classify ((v = computed_value (a))) == -1)
117 || (tarval_classify ((v = computed_value (b))) == -1)) {
122 case iro_Eor: if (ta && tb) { res = tarval_eor (ta, tb); } break;
123 case iro_Not: if(ta) { /*res = tarval_not (ta)*/; } break;
124 case iro_Shl: if (ta && tb) { res = tarval_shl (ta, tb); } break;
125 case iro_Shr: if (ta && tb) { res = tarval_shr (ta, tb); } break;
126 case iro_Shrs: if(ta && tb) { /*res = tarval_shrs (ta, tb)*/; } break;
127 case iro_Rot: if(ta && tb) { /*res = tarval_rot (ta, tb)*/; } break;
128 case iro_Conv: if (ta) { res = tarval_convert_to (ta, get_irn_mode (n)); }
134 a = get_Proj_pred(n);
135 /* Optimize Cmp nodes.
136 This performs a first step of unreachable code elimination.
137 Proj can not be computed, but folding a Cmp above the Proj here is
138 not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
140 There are several case where we can evaluate a Cmp node:
141 1. The nodes compared are both the same. If we compare for
142 equal, this will return true, else it will return false.
143 This step relies on cse.
144 2. The predecessors of Cmp are target values. We can evaluate
146 3. The predecessors are Allocs or void* constants. Allocs never
147 return NULL, they raise an exception. Therefore we can predict
149 if (get_irn_op(a) == op_Cmp) {
150 aa = get_Cmp_left(a);
151 ab = get_Cmp_right(a);
152 if (aa == ab) { /* 1.: */
153 /* This is a tric with the bits used for encoding the Cmp
154 Proj numbers, the following statement is not the same:
155 res = tarval_from_long (mode_b, (get_Proj_proj(n) == Eq)): */
156 res = tarval_from_long (mode_b, (get_Proj_proj(n) & irpn_Eq));
158 tarval *taa = computed_value (aa);
159 tarval *tab = computed_value (ab);
160 if (taa && tab) { /* 2.: */
161 /* strange checks... */
162 ir_pncmp flags = tarval_comp (taa, tab);
163 if (flags != irpn_False) {
164 res = tarval_from_long (mode_b, get_Proj_proj(n) & flags);
166 } else { /* check for 3.: */
167 ir_node *aaa = skip_nop(skip_Proj(aa));
168 ir_node *aba = skip_nop(skip_Proj(ab));
169 if ( ( (/* aa is ProjP and aaa is Alloc */
170 (get_irn_op(aa) == op_Proj)
171 && (get_irn_mode(aa) == mode_p)
172 && (get_irn_op(aaa) == op_Alloc))
173 && ( (/* ab is constant void */
174 (get_irn_op(ab) == op_Const)
175 && (get_irn_mode(ab) == mode_p)
176 && (get_Const_tarval(ab) == tarval_p_void))
177 || (/* ab is other Alloc */
178 (get_irn_op(ab) == op_Proj)
179 && (get_irn_mode(ab) == mode_p)
180 && (get_irn_op(aba) == op_Alloc)
182 || (/* aa is void and aba is Alloc */
183 (get_irn_op(aa) == op_Const)
184 && (get_irn_mode(aa) == mode_p)
185 && (get_Const_tarval(aa) == tarval_p_void)
186 && (get_irn_op(ab) == op_Proj)
187 && (get_irn_mode(ab) == mode_p)
188 && (get_irn_op(aba) == op_Alloc)))
190 res = tarval_from_long (mode_b, get_Proj_proj(n) & irpn_Ne);
194 /* printf(" # comp_val: Proj node, not optimized\n"); */
206 /* returns 1 if the a and b are pointers to different locations. */
208 different_identity (ir_node *a, ir_node *b)
210 assert (get_irn_mode (a) == mode_p
211 && get_irn_mode (b) == mode_p);
213 if (get_irn_op (a) == op_Proj && get_irn_op(b) == op_Proj) {
214 ir_node *a1 = get_Proj_pred (a);
215 ir_node *b1 = get_Proj_pred (b);
216 if (a1 != b1 && get_irn_op (a1) == op_Alloc
217 && get_irn_op (b1) == op_Alloc)
224 /* equivalent_node returns a node equivalent to N. It skips all nodes that
225 perform no actual computation, as, e.g., the Id nodes. It does not create
226 new nodes. It is therefore safe to free N if the node returned is not N.
227 If a node returns a Tuple we can not just skip it. If the size of the
228 in array fits, we transform n into a tuple (e.g., Div). */
230 equivalent_node (ir_node *n)
233 ir_node *a = NULL; /* to shutup gcc */
234 ir_node *b = NULL; /* to shutup gcc */
235 ir_node *c = NULL; /* to shutup gcc */
237 ins = get_irn_arity (n);
239 /* get the operands we will work on */
241 a = get_binop_left(n);
242 b = get_binop_right(n);
243 } else if (is_unop(n)) {
248 /* skip unnecessary nodes. */
249 switch (get_irn_opcode (n)) {
252 /* The Block constructor does not call optimize, therefore
253 dead blocks are not removed without an extra optimizing
254 pass through the graph. */
255 assert(get_Block_matured(n));
257 /* a single entry Region following a single exit Region can be merged */
258 /* !!! Beware, all Phi-nodes of n must have been optimized away.
259 This is true, as the block is matured before optimize is called. */
260 if (get_Block_n_cfgpreds(n) == 1
261 && get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp) {
262 n = get_nodes_Block(get_Block_cfgpred(n, 0));
263 } else if (n != current_ir_graph->start_block) {
266 /* If all inputs are dead, this block is dead too, except if it is
267 the start block. This is a step of unreachable code elimination */
269 for (i = 0; i < get_Block_n_cfgpreds(n); i++) {
270 if (!is_Bad(get_Block_cfgpred(n, i))) break;
272 if (i == get_Block_n_cfgpreds(n))
278 case iro_Jmp: /* GL: ??? Why not same for op_Raise?? */
279 /* unreachable code elimination */
280 if (is_Bad(get_nodes_Block(n))) n = new_Bad();
282 /* GL: Why isn't there a case that checks whether input ot Cond is
283 constant true or false? For unreachable code elimination
284 is this done in Proj? It's not here as it generates a new node,
285 a Jmp. It is in transform_node. *
289 /* remove stuff as x+0, x*1 x&true ... constant expression evaluation */
290 case iro_Or: if (a == b) {n = a; break;}
295 /* After running compute_node there is only one constant predecessor.
296 Find this predecessors value and remember the other node: */
297 if ((tv = computed_value (a))) {
299 } else if ((tv = computed_value (b))) {
303 /* If this predecessors constant value is zero, the operation is
304 unnecessary. Remove it: */
305 if (tarval_classify (tv) == 0) {
315 /* these operations are not commutative. Test only one predecessor. */
316 if (tarval_classify (computed_value (b)) == 0) {
318 /* Test if b > #bits of a ==> return 0 / divide b by #bits
319 --> transform node? */
322 case iro_Not: /* NotNot x == x */
323 case iro_Minus: /* --x == x */ /* ??? Is this possible or can --x raise an
324 out of bounds exception if min =! max? */
325 if (get_irn_op(get_unop_op(n)) == get_irn_op(n))
326 n = get_unop_op(get_unop_op(n));
329 /* Mul is commutative and has again an other neutral element. */
330 if (tarval_classify (computed_value (a)) == 1) {
332 } else if (tarval_classify (computed_value (b)) == 1) {
337 /* Div is not commutative. */
338 if (tarval_classify (computed_value (b)) == 1) { /* div(x, 1) == x */
339 /* Turn Div into a tuple (mem, bad, a) */
340 ir_node *mem = get_Div_mem(n);
341 turn_into_tuple(n, 3);
342 set_Tuple_pred(n, 0, mem);
343 set_Tuple_pred(n, 1, new_Bad());
344 set_Tuple_pred(n, 2, a);
347 /* GL: Why are they skipped? DivMod allocates new nodes --> it's
348 teated in transform node.
349 case iro_Mod, Quot, DivMod
353 /* And has it's own neutral element */
354 else if (tarval_classify (computed_value (a)) == -1) {
356 } else if (tarval_classify (computed_value (b)) == -1) {
361 if (get_irn_mode(n) == get_irn_mode(a)) { /* No Conv necessary */
363 } else if (get_irn_mode(n) == mode_b) {
364 if (get_irn_op(a) == op_Conv &&
365 get_irn_mode (get_Conv_op(a)) == mode_b) {
366 n = get_Conv_op(a); /* Convb(Conv*(xxxb(...))) == xxxb(...) */
373 /* Several optimizations:
374 - no Phi in start block.
375 - remove Id operators that are inputs to Phi
376 - fold Phi-nodes, iff they have only one predecessor except
380 ir_node *block = NULL; /* to shutup gcc */
381 ir_node *first_val = NULL; /* to shutup gcc */
382 ir_node *scnd_val = NULL; /* to shutup gcc */
384 n_preds = get_Phi_n_preds(n);
386 block = get_nodes_Block(n);
387 assert(get_irn_op (block) == op_Block);
389 /* there should be no Phi nodes in the Start region. */
390 if (block == current_ir_graph->start_block) {
395 if (n_preds == 0) { /* Phi of dead Region without predecessors. */
396 /* GL: why not return new_Bad? */
401 /* first we test for a special case: */
402 /* Confirm is a special node fixing additional information for a
403 value that is known at a certain point. This is useful for
404 dataflow analysis. */
406 ir_node *a = follow_Id (get_Phi_pred(n, 0));
407 ir_node *b = follow_Id (get_Phi_pred(n, 1));
408 if ( (get_irn_op(a) == op_Confirm)
409 && (get_irn_op(b) == op_Confirm)
410 && (follow_Id (get_irn_n(a, 0)) == follow_Id(get_irn_n(b, 0)))
411 && (get_irn_n(a, 1) == get_irn_n (b, 1))
412 && (a->data.num == (~b->data.num & irpn_True) )) {
413 n = follow_Id (get_irn_n(a, 0));
419 /* Find first non-self-referencing input */
420 for (i = 0; i < n_preds; ++i) {
421 first_val = follow_Id(get_Phi_pred(n, i));
423 set_Phi_pred(n, i, first_val);
424 if ( (first_val != n) /* not self pointer */
425 && (get_irn_op(first_val) != op_Bad) /* value not dead */
426 && !(is_Bad (get_Block_cfgpred(block, i))) ) { /* not dead control flow */
427 break; /* then found first value. */
431 /* A totally Bad or self-referencing Phi */
432 if (i > n_preds) { n = new_Bad(); break; }
436 /* follow_Id () for rest of inputs, determine if any of these
437 are non-self-referencing */
438 while (++i < n_preds) {
439 scnd_val = follow_Id(get_Phi_pred(n, i));
441 set_Phi_pred(n, i, scnd_val);
443 && (scnd_val != first_val)
444 && (get_irn_op(scnd_val) != op_Bad)
445 && !(is_Bad (get_Block_cfgpred(block, i))) ) {
450 /* Fold, if no multiple distinct non-self-referencing inputs */
454 /* skip the remaining Ids. */
455 while (++i < n_preds)
456 set_Phi_pred(n, i, follow_Id(get_Phi_pred(n, i)));
463 a = skip_Proj(get_Load_mem(n));
464 b = skip_Proj(get_Load_ptr(n));
466 if (get_irn_op(a) == op_Store) {
467 if ( different_identity (b, get_Store_ptr(a))) {
468 /* load and store use different pointers, therefore load
469 needs not take store's memory but the state before. */
470 set_Load_mem (n, get_Store_mem(a));
471 } else if (( 0 /* ???didn't get cryptic test that returns 0 */ )) {
477 /* remove unnecessary store. */
479 a = skip_Proj(get_Store_mem(n));
480 b = get_Store_ptr(n);
481 c = skip_Proj(get_Store_value(n));
483 if (get_irn_op(a) == op_Store
484 && get_Store_ptr(a) == b
485 && skip_Proj(get_Store_value(a)) == c) {
486 /* We have twice exactly the same store -- a write after write. */
488 } else if (get_irn_op(c) == op_Load
489 && (a == c || skip_Proj(get_Load_mem(c)) == a)
490 && get_Load_ptr(c) == b )
491 /* !!!??? and a cryptic test */ {
492 /* We just loaded the value from the same memory, i.e., the store
493 doesn't change the memory -- a write after read. */
494 turn_into_tuple(n, 2);
495 set_Tuple_pred(n, 0, a);
496 set_Tuple_pred(n, 1, new_Bad());
503 a = get_Proj_pred(n);
505 if ( get_irn_op(a) == op_Tuple) {
506 /* Remove the Tuple/Proj combination. */
507 if ( get_Proj_proj(n) <= get_Tuple_n_preds(a) ) {
508 n = get_Tuple_pred(a, get_Proj_proj(n));
510 assert(0); /* This should not happen?! (GL added this assert) */
513 } else if (get_irn_mode(n) == mode_X &&
514 is_Bad(get_nodes_Block(n))) {
515 /* Remove dead control flow. */
529 } /* end equivalent_node() */
532 /* tries several [inplace] [optimizing] transformations and returns a
533 equivalent node. The difference to equivalent_node is that these
534 transformations _do_ generate new nodes, and thus the old node must
535 not be freed even if the equivalent node isn't the old one. */
537 transform_node (ir_node *n)
540 ir_node *a = NULL, *b;
543 switch (get_irn_opcode(n)) {
549 a = get_DivMod_left(n);
550 b = get_DivMod_right(n);
551 mode = get_irn_mode(a);
553 if (!( mode_is_int(get_irn_mode(a))
554 && mode_is_int(get_irn_mode(b))))
558 a = new_Const (mode, tarval_from_long (mode, 1));
559 b = new_Const (mode, tarval_from_long (mode, 0));
566 if (tarval_classify(tb) == 1) {
567 b = new_Const (mode, tarval_from_long (mode, 0));
571 resa = tarval_div (ta, tb);
572 if (!resa) break; /* Causes exception!!! Model by replacing through
573 Jmp for X result!? */
574 resb = tarval_mod (ta, tb);
575 if (!resb) break; /* Causes exception! */
576 a = new_Const (mode, resa);
577 b = new_Const (mode, resb);
580 } else if (tarval_classify (ta) == 0) {
585 if (evaluated) { /* replace by tuple */
586 ir_node *mem = get_DivMod_mem(n);
587 turn_into_tuple(n, 4);
588 set_Tuple_pred(n, 0, mem);
589 set_Tuple_pred(n, 1, new_Bad()); /* no exception */
590 set_Tuple_pred(n, 2, a);
591 set_Tuple_pred(n, 3, b);
592 assert(get_nodes_Block(n));
598 /* Replace the Cond by a Jmp if it branches on a constant
602 a = get_Cond_selector(n);
605 if (ta && (get_irn_mode(a) == mode_b)) {
606 /* It's a boolean Cond, branching on a boolean constant.
607 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
608 jmp = new_r_Jmp(current_ir_graph, get_nodes_Block(n));
609 turn_into_tuple(n, 2);
610 if (tv_val_b(ta) == 1) /* GL: I hope this returns 1 if true */ {
611 set_Tuple_pred(n, 0, new_Bad());
612 set_Tuple_pred(n, 1, jmp);
614 set_Tuple_pred(n, 0, jmp);
615 set_Tuple_pred(n, 1, new_Bad());
617 } else if (ta && (get_irn_mode(a) == mode_I)) {
618 /* I don't want to allow Tuples smaller than the biggest Proj.
619 Also this tuple might get really big...
620 I generate the Jmp here, and remember it in link. Link is used
621 when optimizing Proj. */
622 set_irn_link(n, new_r_Jmp(current_ir_graph, get_nodes_Block(n)));
623 } else if ( ((get_irn_op(get_Cond_selector(n)) == op_Eor)
624 /* || (get_irn_op(get_Cond_selector(a)) == op_Not)*/)
625 && (get_irn_mode(get_Cond_selector(n)) == mode_b)
626 && (tarval_classify(computed_value(get_Eor_right(a))) == 1)) {
627 /* The Eor is a negate. Generate a new Cond without the negate,
628 simulate the negate by exchanging the results. */
629 set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_Block(n),
636 a = get_Proj_pred(n);
638 if ( (get_irn_op(a) == op_Cond)
640 && get_irn_op(get_irn_link(a)) == op_Cond) {
641 /* Use the better Cond if the Proj projs from a Cond which get's
642 its result from an Eor/Not. */
643 assert ( ((get_irn_op(get_Cond_selector(a)) == op_Eor)
644 /* || (get_irn_op(get_Cond_selector(a)) == op_Not)*/)
645 && (get_irn_mode(get_Cond_selector(a)) == mode_b)
646 && (get_irn_op(get_irn_link(a)) == op_Cond)
647 && (get_Cond_selector(get_irn_link(a)) ==
648 get_Eor_left(get_Cond_selector(a))));
649 set_Proj_pred(n, get_irn_link(a));
650 if (get_Proj_proj(n) == 0)
654 } else if ( (get_irn_op(a) == op_Cond)
655 && (get_irn_mode(get_Cond_selector(a)) == mode_I)
657 /* The Cond is a Switch on a Constant */
658 if (get_Proj_proj(n) == tv_val_CHIL(value_of(a))) {
659 /* The always taken branch, reuse the existing Jmp. */
660 if (!get_irn_link(a)) /* well, if it exists ;-> */
661 set_irn_link(a, new_r_Jmp(current_ir_graph, get_nodes_Block(n)));
662 assert(get_irn_op(get_irn_link(a)) == op_Jmp);
665 /* a never taken branch */
671 case iro_Eor: { /* @@@ not tested as boolean Eor not allowed any more. */
673 b = get_Eor_right(n);
675 if ( (get_irn_mode(n) == mode_b)
676 && (get_irn_op(a) == op_Proj)
677 && (get_irn_mode(a) == mode_b)
678 && (tarval_classify (computed_value (b)) == 1)
679 && (get_irn_op(get_Proj_pred(a)) == op_Cmp))
680 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
681 n = new_r_Proj(current_ir_graph, get_nodes_Block(n), get_Proj_pred(a),
682 mode_b, get_negated_pnc(get_Proj_proj(a)));
683 else if ( (get_irn_mode(n) == mode_b)
684 && (tarval_classify (computed_value (b)) == 1))
685 /* The Eor is a Not. Replace it by a Not. */
686 /* ????!!!Extend to bitfield 1111111. */
687 n = new_r_Not(current_ir_graph, get_nodes_Block(n), a, mode_b);
690 case iro_Not: { /* @@@ not tested as boolean Eor not allowed any more. */
693 if ( (get_irn_mode(n) == mode_b)
694 && (get_irn_op(a) == op_Proj)
695 && (get_irn_mode(a) == mode_b)
696 && (get_irn_op(get_Proj_pred(a)) == op_Cmp))
697 /* We negate a Cmp. The Cmp has the negated result anyways! */
698 n = new_r_Proj(current_ir_graph, get_nodes_Block(n), get_Proj_pred(a),
699 mode_b, get_negated_pnc(get_Proj_proj(a)));
707 /***************** Common Subexpression Elimination *****************/
709 /* Compare function for two nodes in the hash table. Gets two */
710 /* nodes as parameters. */
711 /* @@@ a+b != b+a ? */
713 vt_cmp (const void *elt, const void *key)
721 if (a == b) return 0;
723 if ((get_irn_op(a) != get_irn_op(b)) ||
724 (get_irn_mode(a) != get_irn_mode(b))) return 1;
726 /* compare if a's in and b's in are equal */
727 /* GL: we optimize only nodes with in arrays of fixed sizes.
728 if (get_irn_arity (a) != -2) {
729 ins = get_irn_arity (a);
730 if (ins != get_irn_arity (b)) return 1;
731 ain = get_irn_in (a);
732 bin = get_irn_in (b);
735 if (get_irn_arity (a) != get_irn_arity(b))
738 /* compare a->in[0..ins] with b->in[0..ins], i.e., include the block. */
739 /* do if (*ain++ != *bin++) return 1; while (ins--); */
740 for (i = -1; i < get_irn_arity(a); i++)
741 if (get_irn_n(a, i) != get_irn_n(b, i))
745 switch (get_irn_opcode(a)) {
747 return get_irn_const_attr (a) != get_irn_const_attr (b);
749 return get_irn_proj_attr (a) != get_irn_proj_attr (b);
751 return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where)
752 || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type);
754 return (get_irn_free_attr(a) != get_irn_free_attr(b));
756 return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num)
757 || (get_irn_symconst_attr(a).tori.typ != get_irn_symconst_attr(b).tori.typ);
759 return (get_irn_call_attr(a)->kind != get_irn_call_attr(b)->kind)
760 || (get_irn_call_attr(a)->arity != get_irn_call_attr(b)->arity);
762 return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind)
763 || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name)
764 || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner)
765 || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name)
766 || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type)
767 || (get_irn_sel_attr(a).ltyp != get_irn_sel_attr(b).ltyp);
769 return get_irn_phi_attr (a) != get_irn_phi_attr (b);
777 ir_node_hash (ir_node *node)
782 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
783 h = get_irn_arity(node);
785 /* consider all in nodes... except the block. */
786 for (i = 0; i < get_irn_arity(node); i++) {
787 h = 9*h + (unsigned long)get_irn_n(node, i);
791 h = 9*h + (unsigned long) get_irn_mode (node);
793 h = 9*h + (unsigned long) get_irn_op (node);
799 new_identities (void)
801 return new_pset (vt_cmp, TUNE_NIR_NODES);
805 del_identities (pset *value_table)
807 del_pset (value_table);
810 /* Return the canonical node computing the same value as n.
811 Looks up the node in a hash table. */
812 static inline ir_node *
813 identify (pset *value_table, ir_node *n)
817 if (!value_table) return n;
819 switch (get_irn_opcode (n)) {
826 /* for commutative operators perform a OP b == b OP a */
827 if (get_binop_left(n) > get_binop_right(n)) {
828 ir_node *h = get_binop_left(n);
829 set_binop_left(n, get_binop_right(n));
830 set_binop_right(n, h);
836 o = pset_find (value_table, n, ir_node_hash (n));
842 /* Return the canonical node computing the same value as n.
843 Looks up the node in a hash table, enters it in the table
844 if it isn't there yet. */
846 identify_remember (pset *value_table, ir_node *node)
850 if (!value_table) return node;
852 /* lookup or insert in hash table with given hash key. */
853 o = pset_insert (value_table, node, ir_node_hash (node));
855 if (o == node) return node;
860 /* garbage in, garbage out. If a node has a dead input, i.e., the
861 Bad node is input to the node, return the Bad node. */
862 static inline ir_node *
866 ir_op* op = get_irn_op(node);
868 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
869 blocks predecessors is dead. */
870 if ( op != op_Block && op != op_Phi && op != op_Tuple) {
871 for (i = -1; i < get_irn_arity(node); i++) {
872 if (is_Bad(get_irn_n(node, i))) {
882 /* These optimizations deallocate nodes from the obstack.
883 It can only be called if it is guaranteed that no other nodes
884 reference this one, i.e., right after construction of a node. */
886 optimize (ir_node *n)
891 if (!get_optimize()) return NULL;
893 /* if not optimize return n */
895 printf(" attention: empty node!!! \n");
899 /* constant expression evaluation / constant folding */
900 if (get_opt_constant_folding()) {
901 /* constants can not be evaluated */
902 if (get_irn_op(n) != op_Const) {
903 /* try to evaluate */
904 tv = computed_value (n);
906 /* evaluation was succesful -- replace the node. */
907 obstack_free (current_ir_graph->obst, n);
908 return new_Const (get_tv_mode (tv), tv);
909 /* xprintf("* optimize: computed node %I\n", n->op->name); */
913 /* remove unnecessary nodes */
914 if (get_opt_constant_folding() || get_irn_op(n) == op_Phi)
915 n = equivalent_node (n);
917 /** common subexpression elimination **/
918 /* Checks whether n is already available. */
919 /* The block input is used to distinguish different subexpressions. Right
920 now all nodes are pinned to blocks, i.e., the cse only finds common
921 subexpressions within a block. */
924 n = identify (current_ir_graph->value_table, n);
926 /* identify found a cse, so deallocate the old node. */
928 obstack_free (current_ir_graph->obst, old_n);
929 /* The AmRoq fiasco returns n here. Martin's version doesn't. */
932 /* Some more constant expression evaluation that does not allow to
934 if (get_opt_constant_folding())
935 n = transform_node (n);
937 /* Remove nodes with dead (Bad) input. */
939 /* Now we can verify the node, as it has no dead inputs any more. */
942 /* Now we have a legal, useful node. Enter it in hash table for cse */
944 /* aborts ??! set/pset can not handle several hash tables??!
945 No, suddenly it works. */
946 n = identify_remember (current_ir_graph->value_table, n);
949 #if 0 /* GL: what's the use of this?? */
950 if ((current_ir_graph->state & irgs_building) && IR_KEEP_ALIVE (n)) {
951 assert (~current_ir_graph->state & irgs_keep_alives_in_arr);
952 pdeq_putr (current_ir_graph->keep.living, n);
959 /* These optimizations never deallocate nodes. This can cause dead
960 nodes lying on the obstack. Remove these by a dead node elimination,
961 i.e., a copying garbage collection. */
963 optimize_in_place (ir_node *n)
969 /* if not optimize return n */
971 /* Here this is possible. Why? */
975 /* constant expression evaluation / constant folding */
976 if (get_opt_constant_folding()) {
977 /* constants can not be evaluated */
978 if (get_irn_op(n) != op_Const) {
979 /* try to evaluate */
980 tv = computed_value (n);
982 /* evaluation was succesful -- replace the node. */
983 return new_Const (get_tv_mode (tv), tv);
984 /* xprintf("* optimize: computed node %I\n", n->op->name);*/
988 /* remove unnecessary nodes */
989 if (get_opt_constant_folding() || get_irn_op(n) == op_Phi)
990 n = equivalent_node (n);
992 /** common subexpression elimination **/
993 /* Checks whether n is already available. */
994 /* The block input is used to distinguish different subexpressions. Right
995 now all nodes are pinned to blocks, i.e., the cse only finds common
996 subexpressions within a block. */
999 n = identify (current_ir_graph->value_table, n);
1001 /* identify found a cse, so deallocate the old node. */
1003 /* The AmRoq fiasco returns n here. Martin's version doesn't. */
1007 /* Some more constant expression evaluation. */
1008 if (get_opt_constant_folding())
1009 n = transform_node (n);
1012 /* Remove nodes with dead (Bad) input. */
1014 /* Now we can verify the node, as it has no dead inputs any more. */
1017 /* Now we have a legal, useful node. Enter it in hash table for cse */
1018 if (get_opt_cse()) {
1019 /* aborts ??! set/pset can not handle several hash tables??!
1020 No, suddenly it works. */
1021 n = identify_remember (current_ir_graph->value_table, n);