2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Tail-recursion call optimization.
24 * @author Michael Beck
35 #include "iroptimize.h"
36 #include "scalar_replace.h"
43 #include "irgraph_t.h"
51 DEBUG_ONLY(static firm_dbg_module_t *dbg);
54 * the environment for collecting data
56 typedef struct _collect_t {
57 ir_node *proj_X; /**< initial exec proj */
58 ir_node *block; /**< old first block */
59 int blk_idx; /**< cfgpred index of the initial exec in block */
60 ir_node *proj_m; /**< memory from start proj's */
61 ir_node *proj_data; /**< linked list of all parameter access proj's */
65 * walker for collecting data, fills a collect_t environment
67 static void collect_data(ir_node *node, void *env) {
68 collect_t *data = env;
72 switch (get_irn_opcode(node)) {
74 pred = get_Proj_pred(node);
76 op = get_irn_op(pred);
78 ir_node *start = get_Proj_pred(pred);
80 if (is_Start(start)) {
81 if (get_Proj_proj(pred) == pn_Start_T_args) {
82 /* found Proj(ProjT(Start)) */
83 set_irn_link(node, data->proj_data);
84 data->proj_data = node;
87 } else if (op == op_Start) {
88 if (get_Proj_proj(node) == pn_Start_X_initial_exec) {
89 /* found ProjX(Start) */
95 int i, n_pred = get_Block_n_cfgpreds(node);
98 * the first block has the initial exec as cfg predecessor
100 if (node != get_irg_start_block(current_ir_graph)) {
101 for (i = 0; i < n_pred; ++i) {
102 if (get_Block_cfgpred(node, i) == data->proj_X) {
116 typedef enum tail_rec_variants {
117 TR_DIRECT, /**< direct return value, i.e. return func(). */
118 TR_ADD, /**< additive return value, i.e. return x +/- func() */
119 TR_MUL, /**< multiplicative return value, i.e. return x * func() or return -func() */
120 TR_BAD, /**< any other transformation */
121 TR_UNKNOWN /**< during construction */
124 typedef struct tr_env {
125 int n_tail_calls; /**< number of tail calls found */
126 int n_ress; /**< number of return values */
127 tail_rec_variants *variants; /**< return value variants */
128 ir_node *rets; /**< list of returns that can be transformed */
133 * do the graph reconstruction for tail-recursion elimination
135 * @param irg the graph that will reconstructed
136 * @param rets linked list of all rets
137 * @param n_tail_calls number of tail-recursion calls
139 static void do_opt_tail_rec(ir_graph *irg, tr_env *env) {
140 ir_node *end_block = get_irg_end_block(irg);
141 ir_node *block, *jmp, *call, *calls;
144 ir_node ***call_params;
146 int i, j, n_params, n_locs;
148 int rem = get_optimize();
149 ir_entity *ent = get_irg_entity(irg);
150 ir_type *method_tp = get_entity_type(ent);
151 ir_graph *old = current_ir_graph;
153 current_ir_graph = irg;
155 assert(env->n_tail_calls > 0);
157 /* we add new nodes, so the outs are inconsistent */
158 set_irg_outs_inconsistent(irg);
160 /* we add new blocks and change the control flow */
161 set_irg_doms_inconsistent(irg);
162 set_irg_extblk_inconsistent(irg);
164 /* we add a new loop */
165 set_irg_loopinfo_inconsistent(irg);
167 /* calls are removed */
168 set_trouts_inconsistent();
170 /* we must build some new nodes WITHOUT CSE */
173 /* collect needed data */
177 data.proj_m = get_irg_initial_mem(irg);
178 data.proj_data = NULL;
179 irg_walk_graph(irg, NULL, collect_data, &data);
181 /* check number of arguments */
182 call = get_irn_link(end_block);
183 n_params = get_Call_n_params(call);
185 assert(data.proj_X && "Could not find initial exec from Start");
186 assert(data.block && "Could not find first block");
187 assert(data.proj_m && "Could not find initial memory");
188 assert((data.proj_data || n_params == 0) && "Could not find Proj(ProjT(Start)) of non-void function");
190 /* allocate in's for phi and block construction */
191 NEW_ARR_A(ir_node *, in, env->n_tail_calls + 1);
195 /* turn Return's into Jmp's */
196 for (i = 1, p = env->rets; p; p = n) {
197 ir_node *block = get_nodes_block(p);
200 in[i++] = new_r_Jmp(irg, block);
202 // exchange(p, new_r_Bad(irg));
204 /* we might generate an endless loop, so add
205 * the block to the keep-alive list */
206 add_End_keepalive(get_irg_end(irg), block);
209 /* create a new block at start */
210 block = new_r_Block(irg, env->n_tail_calls + 1, in);
211 jmp = new_r_Jmp(irg, block);
213 /* the old first block is now the second one */
214 set_Block_cfgpred(data.block, data.blk_idx, jmp);
216 /* allocate phi's, position 0 contains the memory phi */
217 NEW_ARR_A(ir_node *, phis, n_params + 1);
219 /* build the memory phi */
221 in[i] = new_r_Proj(irg, get_irg_start_block(irg), get_irg_start(irg), mode_M, pn_Start_M);
222 set_irg_initial_mem(irg, in[i]);
225 for (calls = call; calls; calls = get_irn_link(calls)) {
226 in[i] = get_Call_mem(calls);
229 assert(i == env->n_tail_calls + 1);
231 phis[0] = new_r_Phi(irg, block, env->n_tail_calls + 1, in, mode_M);
233 /* build the data Phi's */
239 NEW_ARR_A(ir_node **, call_params, env->n_tail_calls);
241 /* collect all parameters */
242 for (i = 0, calls = call; calls; calls = get_irn_link(calls)) {
243 call_params[i] = get_Call_param_arr(calls);
247 /* build new Proj's and Phi's */
248 args = get_irg_args(irg);
249 args_bl = get_nodes_block(args);
250 for (i = 0; i < n_params; ++i) {
251 ir_mode *mode = get_type_mode(get_method_param_type(method_tp, i));
253 in[0] = new_r_Proj(irg, args_bl, args, mode, i);
254 for (j = 0; j < env->n_tail_calls; ++j)
255 in[j + 1] = call_params[j][i];
257 phis[i + 1] = new_r_Phi(irg, block, env->n_tail_calls + 1, in, mode);
262 * ok, we are here, so we have build and collected all needed Phi's
263 * now exchange all Projs into links to Phi
265 exchange(data.proj_m, phis[0]);
266 for (p = data.proj_data; p; p = n) {
267 long proj = get_Proj_proj(p);
269 assert(0 <= proj && proj < n_params);
271 exchange(p, phis[proj + 1]);
274 /* tail recursion was done, all info is invalid */
275 set_irg_doms_inconsistent(irg);
276 set_irg_outs_inconsistent(irg);
277 set_irg_extblk_inconsistent(irg);
278 set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent);
279 set_trouts_inconsistent();
280 set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
284 /* check if we need new values */
286 for (i = 0; i < env->n_ress; ++i) {
287 if (env->variants[i] != TR_DIRECT)
292 ir_node *bad, *start_block;
296 NEW_ARR_A(ir_node *, in, n_locs);
297 NEW_ARR_A(ir_mode *, modes, n_locs);
298 ssa_cons_start(irg, n_locs);
300 start_block = get_irg_start_block(irg);
301 set_cur_block(start_block);
303 for (i = 0; i < env->n_ress; ++i) {
304 ir_type *tp = get_method_res_type(method_tp, i);
305 ir_mode *mode = get_type_mode(tp);
308 if (env->variants[i] == TR_ADD) {
309 set_value(i, new_Const(mode, get_mode_null(mode)));
310 } else if (env->variants[i] == TR_MUL) {
311 set_value(i, new_Const(mode, get_mode_one(mode)));
314 mature_immBlock(start_block);
316 /* no: we can kill all returns */
317 bad = get_irg_bad(irg);
319 for (p = env->rets; p; p = n) {
320 ir_node *block = get_nodes_block(p);
321 ir_node *call, *mem, *jmp, *tuple;
323 set_cur_block(block);
326 call = skip_Proj(get_Return_mem(p));
327 assert(is_Call(call));
329 mem = get_Call_mem(call);
331 /* create a new jump, free of CSE */
336 for (i = 0; i < env->n_ress; ++i) {
337 if (env->variants[i] != TR_DIRECT) {
338 in[i] = get_value(i, modes[i]);
343 /* create a new tuple for the return values */
344 tuple = new_Tuple(env->n_ress, in);
346 turn_into_tuple(call, pn_Call_max);
347 set_Tuple_pred(call, pn_Call_M, mem);
348 set_Tuple_pred(call, pn_Call_X_regular, jmp);
349 set_Tuple_pred(call, pn_Call_X_except, bad);
350 set_Tuple_pred(call, pn_Call_T_result, tuple);
351 set_Tuple_pred(call, pn_Call_M_except, mem);
352 set_Tuple_pred(call, pn_Call_P_value_res_base, bad);
354 for (i = 0; i < env->n_ress; ++i) {
355 ir_node *res = get_Return_res(p, i);
356 if (env->variants[i] != TR_DIRECT) {
364 /* finally fix all other returns */
365 end_block = get_irg_end_block(irg);
366 for (i = get_Block_n_cfgpreds(end_block) - 1; i >= 0; --i) {
367 ir_node *ret = get_Block_cfgpred(end_block, i);
369 /* search all Returns of a block */
370 if (! is_Return(ret))
373 set_cur_block(get_nodes_block(ret));
374 for (j = 0; j < env->n_ress; ++j) {
375 ir_node *pred = get_Return_res(ret, j);
378 switch (env->variants[j]) {
383 n = get_value(j, modes[j]);
384 n = new_Add(n, pred, modes[j]);
385 set_Return_res(ret, j, n);
389 n = get_value(j, modes[j]);
390 n = new_Mul(n, pred, modes[j]);
391 set_Return_res(ret, j, n);
395 assert(!"unexpected tail recursion variant");
399 ssa_cons_finish(irg);
401 ir_node *bad = get_irg_bad(irg);
403 /* no: we can kill all returns */
404 for (p = env->rets; p; p = n) {
409 current_ir_graph = old;
413 * Check the lifetime of locals in the given graph.
414 * Tail recursion can only be done, if we can prove that
415 * the lifetime of locals end with the recursive call.
416 * We do this by checking that no address of a local variable is
417 * stored or transmitted as an argument to a call.
419 * @return non-zero if it's ok to do tail recursion
421 static int check_lifetime_of_locals(ir_graph *irg) {
422 ir_node *irg_frame, *irg_val_param_base;
425 irg_frame = get_irg_frame(irg);
426 for (i = get_irn_n_outs(irg_frame) - 1; i >= 0; --i) {
427 ir_node *succ = get_irn_out(irg_frame, i);
429 if (is_Sel(succ) && is_address_taken(succ))
433 /* Check if we have compound arguments.
434 For now, we cannot handle them, */
435 irg_val_param_base = get_irg_value_param_base(irg);
436 if (get_irn_n_outs(irg_val_param_base) > 0)
443 * Examine irn and detect the recursion variant.
445 static tail_rec_variants find_variant(ir_node *irn, ir_node *call) {
447 tail_rec_variants va, vb, res;
449 if (skip_Proj(skip_Proj(irn)) == call) {
453 switch (get_irn_opcode(irn)) {
456 a = get_Add_left(irn);
457 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(call)) {
458 /* we are outside, ignore */
461 va = find_variant(a, call);
465 b = get_Add_right(irn);
466 if (get_irn_MacroBlock(b) != get_irn_MacroBlock(call)) {
467 /* we are outside, ignore */
470 vb = find_variant(b, call);
477 else if (va == TR_UNKNOWN)
479 else if (vb == TR_UNKNOWN)
482 /* they are different but none is TR_UNKNOWN -> incompatible */
485 if (res == TR_DIRECT || res == TR_ADD)
491 /* try additive, but return value must be left */
492 a = get_Sub_left(irn);
493 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(call)) {
494 /* we are outside, ignore */
497 va = find_variant(a, call);
501 b = get_Sub_right(irn);
502 if (get_irn_MacroBlock(b) != get_irn_MacroBlock(call)) {
503 /* we are outside, ignore */
506 vb = find_variant(b, call);
507 if (vb != TR_UNKNOWN)
511 if (res == TR_DIRECT || res == TR_ADD)
517 /* try multiplicative */
518 a = get_Mul_left(irn);
519 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(call)) {
520 /* we are outside, ignore */
523 va = find_variant(a, call);
527 b = get_Mul_right(irn);
528 if (get_irn_MacroBlock(b) != get_irn_MacroBlock(call)) {
529 /* we are outside, ignore */
532 vb = find_variant(b, call);
539 else if (va == TR_UNKNOWN)
541 else if (vb == TR_UNKNOWN)
544 /* they are different but none is TR_UNKNOWN -> incompatible */
547 if (res == TR_DIRECT || res == TR_MUL)
553 /* try multiplicative */
554 a = get_Minus_op(irn);
555 res = find_variant(a, call);
556 if (res == TR_DIRECT)
558 if (res == TR_MUL || res == TR_UNKNOWN)
570 * convert simple tail-calls into loops
572 int opt_tail_rec_irg(ir_graph *irg) {
575 int i, n_ress, n_tail_calls = 0;
576 ir_node *rets = NULL;
577 ir_type *mtd_type, *call_type;
580 assure_irg_outs(irg);
582 if (! check_lifetime_of_locals(irg))
586 ent = get_irg_entity(irg);
587 mtd_type = get_entity_type(ent);
588 n_ress = get_method_n_ress(mtd_type);
594 NEW_ARR_A(tail_rec_variants, env.variants, n_ress);
596 for (i = 0; i < n_ress; ++i)
597 env.variants[i] = TR_DIRECT;
601 * This tail recursion optimization works best
602 * if the Returns are normalized.
604 normalize_n_returns(irg);
606 end_block = get_irg_end_block(irg);
607 set_irn_link(end_block, NULL);
609 for (i = get_Block_n_cfgpreds(end_block) - 1; i >= 0; --i) {
610 ir_node *ret = get_Block_cfgpred(end_block, i);
611 ir_node *call, *call_ptr;
615 /* search all Returns of a block */
616 if (! is_Return(ret))
619 /* check, if it's a Return self() */
620 call = skip_Proj(get_Return_mem(ret));
624 /* the call must be in the same block as the return */
625 if (get_nodes_block(call) != get_nodes_block(ret))
628 /* check if it's a recursive call */
629 call_ptr = get_Call_ptr(call);
631 if (! is_Global(call_ptr))
634 ent = get_Global_entity(call_ptr);
635 if (!ent || get_entity_irg(ent) != irg)
639 * Check, that the types match. At least in C
642 mtd_type = get_entity_type(ent);
643 call_type = get_Call_type(call);
645 if (mtd_type != call_type) {
647 * Hmm, the types did not match, bad.
648 * This can happen in C when no prototype is given
649 * or K&R style is used.
652 printf("Warning: Tail recursion fails because of different method and call types:\n");
654 dump_type(call_type);
659 /* ok, mem is routed to a recursive call, check return args */
660 ress = get_Return_res_arr(ret);
661 for (j = get_Return_n_ress(ret) - 1; j >= 0; --j) {
662 tail_rec_variants var = find_variant(ress[j], call);
665 /* cannot be transformed */
668 if (var == TR_DIRECT)
669 var = env.variants[j];
670 else if (env.variants[j] == TR_DIRECT)
671 env.variants[j] = var;
672 if (env.variants[j] != var) {
680 /* here, we have found a call */
681 set_irn_link(call, get_irn_link(end_block));
682 set_irn_link(end_block, call);
685 /* link all returns, we will need this */
686 set_irn_link(ret, rets);
690 /* now, end_block->link contains the list of all tail calls */
691 if (n_tail_calls <= 0)
694 DB((dbg, LEVEL_2, " Performing tail recursion for graph %s and %d Calls\n",
695 get_entity_ld_name(get_irg_entity(irg)), n_tail_calls));
697 hook_tail_rec(irg, n_tail_calls);
699 env.n_tail_calls = n_tail_calls;
701 do_opt_tail_rec(irg, &env);
707 * optimize tail recursion away
709 void opt_tail_recursion(void) {
711 int n_opt_applications = 0;
714 FIRM_DBG_REGISTER(dbg, "firm.opt.tailrec");
716 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
717 irg = get_irp_irg(i);
719 current_ir_graph = irg;
721 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
722 if (opt_tail_rec_irg(irg))
723 ++n_opt_applications;
725 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
728 DB((dbg, LEVEL_1, "Performed tail recursion for %d of %d graphs\n",
729 n_opt_applications, get_irp_n_irgs()));