2 * Copyright (C) 1995-2007 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 * File name: ir/opt/tailrec.c
23 * Purpose: tail-recursion optimization
24 * Author: Michael Beck
28 * Copyright: (c) 2002-2004 Universität Karlsruhe
46 #include "irgraph_t.h"
51 #include "scalar_replace.h"
57 * the environment for collecting data
59 typedef struct _collect_t {
60 ir_node *proj_X; /**< initial exec proj */
61 ir_node *block; /**< old first block */
62 int blk_idx; /**< cfgpred index of the initial exec in block */
63 ir_node *proj_m; /**< memory from start proj's */
64 ir_node *proj_data; /**< linked list of all parameter access proj's */
68 * walker for collecting data, fills a collect_t environment
70 static void collect_data(ir_node *node, void *env)
72 collect_t *data = env;
76 switch (get_irn_opcode(node)) {
78 pred = get_Proj_pred(node);
80 op = get_irn_op(pred);
82 ir_node *start = get_Proj_pred(pred);
84 if (get_irn_op(start) == op_Start) {
85 if (get_Proj_proj(pred) == pn_Start_T_args) {
86 /* found Proj(ProjT(Start)) */
87 set_irn_link(node, data->proj_data);
88 data->proj_data = node;
92 else if (op == op_Start) {
93 if (get_Proj_proj(node) == pn_Start_X_initial_exec) {
94 /* found ProjX(Start) */
100 int i, n_pred = get_Block_n_cfgpreds(node);
103 * the first block has the initial exec as cfg predecessor
105 if (node != get_irg_start_block(current_ir_graph)) {
106 for (i = 0; i < n_pred; ++i) {
107 if (get_Block_cfgpred(node, i) == data->proj_X) {
122 * do the graph reconstruction for tail-recursion elimination
124 * @param irg the graph that will reconstructed
125 * @param rets linked list of all rets
126 * @param n_tail_calls number of tail-recursion calls
128 static void do_opt_tail_rec(ir_graph *irg, ir_node *rets, int n_tail_calls)
130 ir_node *end_block = get_irg_end_block(irg);
131 ir_node *block, *jmp, *call, *calls;
134 ir_node ***call_params;
138 int rem = get_optimize();
139 ir_entity *ent = get_irg_entity(irg);
140 ir_type *method_tp = get_entity_type(ent);
142 assert(n_tail_calls);
144 /* we add new nodes, so the outs are inconsistent */
145 set_irg_outs_inconsistent(irg);
147 /* we add new blocks and change the control flow */
148 set_irg_doms_inconsistent(irg);
149 set_irg_extblk_inconsistent(irg);
151 /* we add a new loop */
152 set_irg_loopinfo_inconsistent(irg);
154 /* calls are removed */
155 set_trouts_inconsistent();
157 /* we must build some new nodes WITHOUT CSE */
160 /* collect needed data */
164 data.proj_m = get_irg_initial_mem(irg);
165 data.proj_data = NULL;
166 irg_walk_graph(irg, NULL, collect_data, &data);
168 /* check number of arguments */
169 call = get_irn_link(end_block);
170 n_params = get_Call_n_params(call);
172 assert(data.proj_X && "Could not find initial exec from Start");
173 assert(data.block && "Could not find first block");
174 assert(data.proj_m && "Could not find initial memory");
175 assert((data.proj_data || n_params == 0) && "Could not find Proj(ProjT(Start)) of non-void function");
177 /* allocate in's for phi and block construction */
178 NEW_ARR_A(ir_node *, in, n_tail_calls + 1);
182 /* turn Return's into Jmp's */
183 for (i = 1, p = rets; p; p = n) {
184 ir_node *block = get_nodes_block(p);
187 in[i++] = new_r_Jmp(irg, block);
189 exchange(p, new_r_Bad(irg));
191 /* we might generate an endless loop, so add
192 * the block to the keep-alive list */
193 add_End_keepalive(get_irg_end(irg), block);
196 /* create a new block at start */
197 block = new_r_Block(irg, n_tail_calls + 1, in);
198 jmp = new_r_Jmp(irg, block);
200 /* the old first block is now the second one */
201 set_Block_cfgpred(data.block, data.blk_idx, jmp);
203 /* allocate phi's, position 0 contains the memory phi */
204 NEW_ARR_A(ir_node *, phis, n_params + 1);
206 /* build the memory phi */
208 in[i] = new_r_Proj(irg, get_irg_start_block(irg), get_irg_start(irg), mode_M, pn_Start_M);
209 set_irg_initial_mem(irg, in[i]);
212 for (calls = call; calls; calls = get_irn_link(calls)) {
213 in[i] = get_Call_mem(calls);
216 assert(i == n_tail_calls + 1);
218 phis[0] = new_r_Phi(irg, block, n_tail_calls + 1, in, mode_M);
220 /* build the data Phi's */
226 NEW_ARR_A(ir_node **, call_params, n_tail_calls);
228 /* collect all parameters */
229 for (i = 0, calls = call; calls; calls = get_irn_link(calls)) {
230 call_params[i] = get_Call_param_arr(calls);
234 /* build new Proj's and Phi's */
235 args = get_irg_args(irg);
236 args_bl = get_nodes_block(args);
237 for (i = 0; i < n_params; ++i) {
238 ir_mode *mode = get_type_mode(get_method_param_type(method_tp, i));
240 in[0] = new_r_Proj(irg, args_bl, args, mode, i);
241 for (j = 0; j < n_tail_calls; ++j)
242 in[j + 1] = call_params[j][i];
244 phis[i + 1] = new_r_Phi(irg, block, n_tail_calls + 1, in, mode);
249 * ok, we are here, so we have build and collected all needed Phi's
250 * now exchange all Projs into links to Phi
252 for (p = data.proj_m; p; p = n) {
254 exchange(p, phis[0]);
256 for (p = data.proj_data; p; p = n) {
257 long proj = get_Proj_proj(p);
259 assert(0 <= proj && proj < n_params);
261 exchange(p, phis[proj + 1]);
264 /* tail recursion was done, all info is invalid */
265 set_irg_doms_inconsistent(irg);
266 set_irg_outs_inconsistent(irg);
267 set_irg_extblk_inconsistent(irg);
268 set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent);
269 set_trouts_inconsistent();
270 set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
276 * Check the lifetime of locals in the given graph.
277 * Tail recursion can only be done, if we can prove that
278 * the lifetime of locals end with the recursive call.
279 * We do this by checking that no address of a local variable is
280 * stored or transmitted as an argument to a call.
282 * @return non-zero if it's ok to do tail recursion
284 static int check_lifetime_of_locals(ir_graph *irg)
286 ir_node *irg_frame = get_irg_frame(irg);
289 if (get_irg_outs_state(irg) != outs_consistent)
290 compute_irg_outs(irg);
292 for (i = get_irn_n_outs(irg_frame) - 1; i >= 0; --i) {
293 ir_node *succ = get_irn_out(irg_frame, i);
295 if (is_Sel(succ) && is_address_taken(succ))
302 * convert simple tail-calls into loops
304 int opt_tail_rec_irg(ir_graph *irg)
307 int i, n_tail_calls = 0;
308 ir_node *rets = NULL;
309 ir_type *mtd_type, *call_type;
311 if (! get_opt_tail_recursion() || ! get_opt_optimize())
314 if (! check_lifetime_of_locals(irg))
318 * This tail recursion optimization works best
319 * if the Returns are normalized.
321 normalize_n_returns(irg);
323 end_block = get_irg_end_block(irg);
324 set_irn_link(end_block, NULL);
326 for (i = get_Block_n_cfgpreds(end_block) - 1; i >= 0; --i) {
327 ir_node *ret = get_Block_cfgpred(end_block, i);
328 ir_node *call, *call_ptr;
333 /* search all Returns of a block */
334 if (! is_Return(ret))
337 /* check, if it's a Return self() */
338 call = skip_Proj(get_Return_mem(ret));
342 /* check if it's a recursive call */
343 call_ptr = get_Call_ptr(call);
345 if (get_irn_op(call_ptr) != op_SymConst)
348 if (get_SymConst_kind(call_ptr) != symconst_addr_ent)
351 ent = get_SymConst_entity(call_ptr);
352 if (!ent || get_entity_irg(ent) != irg)
355 /* ok, mem is routed to a recursive call, check return args */
356 ress = get_Return_res_arr(ret);
357 for (j = get_Return_n_ress(ret) - 1; j >= 0; --j) {
358 ir_node *irn = skip_Proj(skip_Proj(ress[j]));
361 /* not routed to a call */
369 * Check, that the types match. At least in C
372 mtd_type = get_entity_type(ent);
373 call_type = get_Call_type(call);
375 if (mtd_type != call_type) {
377 * Hmm, the types did not match, bad.
378 * This can happen in C when no prototype is given
379 * or K&R style is used.
382 printf("Warning: Tail recursion fails because of different method and call types:\n");
384 dump_type(call_type);
389 /* here, we have found a call */
390 set_irn_link(call, get_irn_link(end_block));
391 set_irn_link(end_block, call);
394 /* link all returns, we will need this */
395 set_irn_link(ret, rets);
399 /* now, end_block->link contains the list of all tail calls */
403 if (get_opt_tail_recursion_verbose() && get_firm_verbosity() > 1)
404 printf(" Performing tail recursion for graph %s and %d Calls\n",
405 get_entity_ld_name(get_irg_entity(irg)), n_tail_calls);
407 hook_tail_rec(irg, n_tail_calls);
408 do_opt_tail_rec(irg, rets, n_tail_calls);
414 * optimize tail recursion away
416 void opt_tail_recursion(void)
419 int n_opt_applications = 0;
422 if (! get_opt_tail_recursion() || ! get_opt_optimize())
425 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
426 irg = get_irp_irg(i);
428 if (opt_tail_rec_irg(irg))
429 ++n_opt_applications;
432 if (get_opt_tail_recursion_verbose())
433 printf("Performed tail recursion for %d of %d graphs\n", n_opt_applications, get_irp_n_irgs());