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
34 #include "iroptimize.h"
35 #include "scalar_replace.h"
42 #include "irgraph_t.h"
51 * the environment for collecting data
53 typedef struct _collect_t {
54 ir_node *proj_X; /**< initial exec proj */
55 ir_node *block; /**< old first block */
56 int blk_idx; /**< cfgpred index of the initial exec in block */
57 ir_node *proj_m; /**< memory from start proj's */
58 ir_node *proj_data; /**< linked list of all parameter access proj's */
62 * walker for collecting data, fills a collect_t environment
64 static void collect_data(ir_node *node, void *env) {
65 collect_t *data = env;
69 switch (get_irn_opcode(node)) {
71 pred = get_Proj_pred(node);
73 op = get_irn_op(pred);
75 ir_node *start = get_Proj_pred(pred);
77 if (get_irn_op(start) == op_Start) {
78 if (get_Proj_proj(pred) == pn_Start_T_args) {
79 /* found Proj(ProjT(Start)) */
80 set_irn_link(node, data->proj_data);
81 data->proj_data = node;
84 } else if (op == op_Start) {
85 if (get_Proj_proj(node) == pn_Start_X_initial_exec) {
86 /* found ProjX(Start) */
92 int i, n_pred = get_Block_n_cfgpreds(node);
95 * the first block has the initial exec as cfg predecessor
97 if (node != get_irg_start_block(current_ir_graph)) {
98 for (i = 0; i < n_pred; ++i) {
99 if (get_Block_cfgpred(node, i) == data->proj_X) {
114 * do the graph reconstruction for tail-recursion elimination
116 * @param irg the graph that will reconstructed
117 * @param rets linked list of all rets
118 * @param n_tail_calls number of tail-recursion calls
120 static void do_opt_tail_rec(ir_graph *irg, ir_node *rets, int n_tail_calls) {
121 ir_node *end_block = get_irg_end_block(irg);
122 ir_node *block, *jmp, *call, *calls;
125 ir_node ***call_params;
129 int rem = get_optimize();
130 ir_entity *ent = get_irg_entity(irg);
131 ir_type *method_tp = get_entity_type(ent);
133 assert(n_tail_calls);
135 /* we add new nodes, so the outs are inconsistent */
136 set_irg_outs_inconsistent(irg);
138 /* we add new blocks and change the control flow */
139 set_irg_doms_inconsistent(irg);
140 set_irg_extblk_inconsistent(irg);
142 /* we add a new loop */
143 set_irg_loopinfo_inconsistent(irg);
145 /* calls are removed */
146 set_trouts_inconsistent();
148 /* we must build some new nodes WITHOUT CSE */
151 /* collect needed data */
155 data.proj_m = get_irg_initial_mem(irg);
156 data.proj_data = NULL;
157 irg_walk_graph(irg, NULL, collect_data, &data);
159 /* check number of arguments */
160 call = get_irn_link(end_block);
161 n_params = get_Call_n_params(call);
163 assert(data.proj_X && "Could not find initial exec from Start");
164 assert(data.block && "Could not find first block");
165 assert(data.proj_m && "Could not find initial memory");
166 assert((data.proj_data || n_params == 0) && "Could not find Proj(ProjT(Start)) of non-void function");
168 /* allocate in's for phi and block construction */
169 NEW_ARR_A(ir_node *, in, n_tail_calls + 1);
173 /* turn Return's into Jmp's */
174 for (i = 1, p = rets; p; p = n) {
175 ir_node *block = get_nodes_block(p);
178 in[i++] = new_r_Jmp(irg, block);
180 exchange(p, new_r_Bad(irg));
182 /* we might generate an endless loop, so add
183 * the block to the keep-alive list */
184 add_End_keepalive(get_irg_end(irg), block);
187 /* create a new block at start */
188 block = new_r_Block(irg, n_tail_calls + 1, in);
189 jmp = new_r_Jmp(irg, block);
191 /* the old first block is now the second one */
192 set_Block_cfgpred(data.block, data.blk_idx, jmp);
194 /* allocate phi's, position 0 contains the memory phi */
195 NEW_ARR_A(ir_node *, phis, n_params + 1);
197 /* build the memory phi */
199 in[i] = new_r_Proj(irg, get_irg_start_block(irg), get_irg_start(irg), mode_M, pn_Start_M);
200 set_irg_initial_mem(irg, in[i]);
203 for (calls = call; calls; calls = get_irn_link(calls)) {
204 in[i] = get_Call_mem(calls);
207 assert(i == n_tail_calls + 1);
209 phis[0] = new_r_Phi(irg, block, n_tail_calls + 1, in, mode_M);
211 /* build the data Phi's */
217 NEW_ARR_A(ir_node **, call_params, n_tail_calls);
219 /* collect all parameters */
220 for (i = 0, calls = call; calls; calls = get_irn_link(calls)) {
221 call_params[i] = get_Call_param_arr(calls);
225 /* build new Proj's and Phi's */
226 args = get_irg_args(irg);
227 args_bl = get_nodes_block(args);
228 for (i = 0; i < n_params; ++i) {
229 ir_mode *mode = get_type_mode(get_method_param_type(method_tp, i));
231 in[0] = new_r_Proj(irg, args_bl, args, mode, i);
232 for (j = 0; j < n_tail_calls; ++j)
233 in[j + 1] = call_params[j][i];
235 phis[i + 1] = new_r_Phi(irg, block, n_tail_calls + 1, in, mode);
240 * ok, we are here, so we have build and collected all needed Phi's
241 * now exchange all Projs into links to Phi
243 for (p = data.proj_m; p; p = n) {
245 exchange(p, phis[0]);
247 for (p = data.proj_data; p; p = n) {
248 long proj = get_Proj_proj(p);
250 assert(0 <= proj && proj < n_params);
252 exchange(p, phis[proj + 1]);
255 /* tail recursion was done, all info is invalid */
256 set_irg_doms_inconsistent(irg);
257 set_irg_outs_inconsistent(irg);
258 set_irg_extblk_inconsistent(irg);
259 set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent);
260 set_trouts_inconsistent();
261 set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
267 * Check the lifetime of locals in the given graph.
268 * Tail recursion can only be done, if we can prove that
269 * the lifetime of locals end with the recursive call.
270 * We do this by checking that no address of a local variable is
271 * stored or transmitted as an argument to a call.
273 * @return non-zero if it's ok to do tail recursion
275 static int check_lifetime_of_locals(ir_graph *irg) {
276 ir_node *irg_frame = get_irg_frame(irg);
279 if (get_irg_outs_state(irg) != outs_consistent)
280 compute_irg_outs(irg);
282 for (i = get_irn_n_outs(irg_frame) - 1; i >= 0; --i) {
283 ir_node *succ = get_irn_out(irg_frame, i);
285 if (is_Sel(succ) && is_address_taken(succ))
292 * convert simple tail-calls into loops
294 int opt_tail_rec_irg(ir_graph *irg) {
296 int i, n_tail_calls = 0;
297 ir_node *rets = NULL;
298 ir_type *mtd_type, *call_type;
300 if (! get_opt_tail_recursion() || ! get_opt_optimize())
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 if (get_opt_tail_recursion_verbose() && get_firm_verbosity() > 1)
394 printf(" Performing tail recursion for graph %s and %d Calls\n",
395 get_entity_ld_name(get_irg_entity(irg)), n_tail_calls);
397 hook_tail_rec(irg, n_tail_calls);
398 do_opt_tail_rec(irg, rets, n_tail_calls);
404 * optimize tail recursion away
406 void opt_tail_recursion(void) {
408 int n_opt_applications = 0;
411 if (! get_opt_tail_recursion() || ! get_opt_optimize())
414 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
415 irg = get_irp_irg(i);
417 current_ir_graph = irg;
419 if (opt_tail_rec_irg(irg))
420 ++n_opt_applications;
423 if (get_opt_tail_recursion_verbose())
424 printf("Performed tail recursion for %d of %d graphs\n", n_opt_applications, get_irp_n_irgs());