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 * @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 (get_irn_op(start) == op_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) {
117 * do the graph reconstruction for tail-recursion elimination
119 * @param irg the graph that will reconstructed
120 * @param rets linked list of all rets
121 * @param n_tail_calls number of tail-recursion calls
123 static void do_opt_tail_rec(ir_graph *irg, ir_node *rets, int n_tail_calls) {
124 ir_node *end_block = get_irg_end_block(irg);
125 ir_node *block, *jmp, *call, *calls;
128 ir_node ***call_params;
132 int rem = get_optimize();
133 ir_entity *ent = get_irg_entity(irg);
134 ir_type *method_tp = get_entity_type(ent);
136 assert(n_tail_calls);
138 /* we add new nodes, so the outs are inconsistent */
139 set_irg_outs_inconsistent(irg);
141 /* we add new blocks and change the control flow */
142 set_irg_doms_inconsistent(irg);
143 set_irg_extblk_inconsistent(irg);
145 /* we add a new loop */
146 set_irg_loopinfo_inconsistent(irg);
148 /* calls are removed */
149 set_trouts_inconsistent();
151 /* we must build some new nodes WITHOUT CSE */
154 /* collect needed data */
158 data.proj_m = get_irg_initial_mem(irg);
159 data.proj_data = NULL;
160 irg_walk_graph(irg, NULL, collect_data, &data);
162 /* check number of arguments */
163 call = get_irn_link(end_block);
164 n_params = get_Call_n_params(call);
166 assert(data.proj_X && "Could not find initial exec from Start");
167 assert(data.block && "Could not find first block");
168 assert(data.proj_m && "Could not find initial memory");
169 assert((data.proj_data || n_params == 0) && "Could not find Proj(ProjT(Start)) of non-void function");
171 /* allocate in's for phi and block construction */
172 NEW_ARR_A(ir_node *, in, n_tail_calls + 1);
176 /* turn Return's into Jmp's */
177 for (i = 1, p = rets; p; p = n) {
178 ir_node *block = get_nodes_block(p);
181 in[i++] = new_r_Jmp(irg, block);
183 exchange(p, new_r_Bad(irg));
185 /* we might generate an endless loop, so add
186 * the block to the keep-alive list */
187 add_End_keepalive(get_irg_end(irg), block);
190 /* create a new block at start */
191 block = new_r_Block(irg, n_tail_calls + 1, in);
192 jmp = new_r_Jmp(irg, block);
194 /* the old first block is now the second one */
195 set_Block_cfgpred(data.block, data.blk_idx, jmp);
197 /* allocate phi's, position 0 contains the memory phi */
198 NEW_ARR_A(ir_node *, phis, n_params + 1);
200 /* build the memory phi */
202 in[i] = new_r_Proj(irg, get_irg_start_block(irg), get_irg_start(irg), mode_M, pn_Start_M);
203 set_irg_initial_mem(irg, in[i]);
206 for (calls = call; calls; calls = get_irn_link(calls)) {
207 in[i] = get_Call_mem(calls);
210 assert(i == n_tail_calls + 1);
212 phis[0] = new_r_Phi(irg, block, n_tail_calls + 1, in, mode_M);
214 /* build the data Phi's */
220 NEW_ARR_A(ir_node **, call_params, n_tail_calls);
222 /* collect all parameters */
223 for (i = 0, calls = call; calls; calls = get_irn_link(calls)) {
224 call_params[i] = get_Call_param_arr(calls);
228 /* build new Proj's and Phi's */
229 args = get_irg_args(irg);
230 args_bl = get_nodes_block(args);
231 for (i = 0; i < n_params; ++i) {
232 ir_mode *mode = get_type_mode(get_method_param_type(method_tp, i));
234 in[0] = new_r_Proj(irg, args_bl, args, mode, i);
235 for (j = 0; j < n_tail_calls; ++j)
236 in[j + 1] = call_params[j][i];
238 phis[i + 1] = new_r_Phi(irg, block, n_tail_calls + 1, in, mode);
243 * ok, we are here, so we have build and collected all needed Phi's
244 * now exchange all Projs into links to Phi
246 for (p = data.proj_m; p; p = n) {
248 exchange(p, phis[0]);
250 for (p = data.proj_data; p; p = n) {
251 long proj = get_Proj_proj(p);
253 assert(0 <= proj && proj < n_params);
255 exchange(p, phis[proj + 1]);
258 /* tail recursion was done, all info is invalid */
259 set_irg_doms_inconsistent(irg);
260 set_irg_outs_inconsistent(irg);
261 set_irg_extblk_inconsistent(irg);
262 set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent);
263 set_trouts_inconsistent();
264 set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
270 * Check the lifetime of locals in the given graph.
271 * Tail recursion can only be done, if we can prove that
272 * the lifetime of locals end with the recursive call.
273 * We do this by checking that no address of a local variable is
274 * stored or transmitted as an argument to a call.
276 * @return non-zero if it's ok to do tail recursion
278 static int check_lifetime_of_locals(ir_graph *irg) {
279 ir_node *irg_frame = get_irg_frame(irg);
282 if (get_irg_outs_state(irg) != outs_consistent)
283 compute_irg_outs(irg);
285 for (i = get_irn_n_outs(irg_frame) - 1; i >= 0; --i) {
286 ir_node *succ = get_irn_out(irg_frame, i);
288 if (is_Sel(succ) && is_address_taken(succ))
295 * convert simple tail-calls into loops
297 int opt_tail_rec_irg(ir_graph *irg) {
299 int i, n_tail_calls = 0;
300 ir_node *rets = NULL;
301 ir_type *mtd_type, *call_type;
303 if (! check_lifetime_of_locals(irg))
307 * This tail recursion optimization works best
308 * if the Returns are normalized.
310 normalize_n_returns(irg);
312 end_block = get_irg_end_block(irg);
313 set_irn_link(end_block, NULL);
315 for (i = get_Block_n_cfgpreds(end_block) - 1; i >= 0; --i) {
316 ir_node *ret = get_Block_cfgpred(end_block, i);
317 ir_node *call, *call_ptr;
322 /* search all Returns of a block */
323 if (! is_Return(ret))
326 /* check, if it's a Return self() */
327 call = skip_Proj(get_Return_mem(ret));
331 /* the call must be in the same block as the return */
332 if (get_nodes_block(call) != get_nodes_block(ret))
335 /* check if it's a recursive call */
336 call_ptr = get_Call_ptr(call);
338 if (! is_SymConst(call_ptr) || get_SymConst_kind(call_ptr) != symconst_addr_ent)
341 ent = get_SymConst_entity(call_ptr);
342 if (!ent || get_entity_irg(ent) != irg)
345 /* ok, mem is routed to a recursive call, check return args */
346 ress = get_Return_res_arr(ret);
347 for (j = get_Return_n_ress(ret) - 1; j >= 0; --j) {
348 ir_node *irn = skip_Proj(skip_Proj(ress[j]));
351 /* not routed to a call */
359 * Check, that the types match. At least in C
362 mtd_type = get_entity_type(ent);
363 call_type = get_Call_type(call);
365 if (mtd_type != call_type) {
367 * Hmm, the types did not match, bad.
368 * This can happen in C when no prototype is given
369 * or K&R style is used.
372 printf("Warning: Tail recursion fails because of different method and call types:\n");
374 dump_type(call_type);
379 /* here, we have found a call */
380 set_irn_link(call, get_irn_link(end_block));
381 set_irn_link(end_block, call);
384 /* link all returns, we will need this */
385 set_irn_link(ret, rets);
389 /* now, end_block->link contains the list of all tail calls */
393 DB((dbg, LEVEL_2, " Performing tail recursion for graph %s and %d Calls\n",
394 get_entity_ld_name(get_irg_entity(irg)), n_tail_calls));
396 hook_tail_rec(irg, n_tail_calls);
397 do_opt_tail_rec(irg, rets, n_tail_calls);
403 * optimize tail recursion away
405 void opt_tail_recursion(void) {
407 int n_opt_applications = 0;
410 FIRM_DBG_REGISTER(dbg, "firm.opt.tailrec");
412 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
413 irg = get_irp_irg(i);
415 current_ir_graph = irg;
417 if (opt_tail_rec_irg(irg))
418 ++n_opt_applications;
421 DB((dbg, LEVEL_1, "Performed tail recursion for %d of %d graphs\n",
422 n_opt_applications, get_irp_n_irgs()));