3 * File name: ir/opt/tailrec.c
4 * Purpose: tail-recursion optimization
9 * Copyright: (c) 2002-2004 Universität Karlsruhe
10 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
34 #include "irgraph_t.h"
41 * the environment for collecting data
43 typedef struct _collect_t {
44 ir_node *proj_X; /**< initial exec proj */
45 ir_node *block; /**< old first block */
46 int blk_idx; /**< cfgpred index of the initial exec in block */
47 ir_node *proj_m; /**< linked list of memory from start proj's */
48 ir_node *proj_data; /**< linked list of all parameter access proj's */
52 * walker for collecting data, fills a collect_t environment
54 static void collect_data(ir_node *node, void *env)
56 collect_t *data = env;
60 switch (get_irn_opcode(node)) {
62 pred = get_Proj_pred(node);
64 op = get_irn_op(pred);
66 ir_node *start = get_Proj_pred(pred);
68 if (get_irn_op(start) == op_Start) {
69 if (get_Proj_proj(pred) == pn_Start_T_args) {
70 /* found Proj(ProjT(Start)) */
71 set_irn_link(node, data->proj_data);
72 data->proj_data = node;
76 else if (op == op_Start) {
77 switch (get_Proj_proj(node)) {
79 /* found ProjM(Start) */
80 set_irn_link(node, data->proj_m);
83 case pn_Start_X_initial_exec:
84 /* found ProjX(Start) */
93 int i, n_pred = get_Block_n_cfgpreds(node);
96 * the first block has the initial exec as cfg predecessor
98 if (node != current_ir_graph->start_block) {
99 for (i = 0; i < n_pred; ++i) {
100 if (get_Block_cfgpred(node, i) == data->proj_X) {
115 * do the graph reconstruction for tail-recursion elimination
117 * @param irg the graph that will reconstructed
118 * @param rets linked list of all rets
119 * @param n_tail_calls number of tail-recursion calls
121 static void do_opt_tail_rec(ir_graph *irg, ir_node *rets, int n_tail_calls)
123 ir_graph *rem_irg = current_ir_graph;
124 ir_node *end_block = irg->end_block;
125 ir_node *block, *jmp, *call, *calls;
128 ir_node ***call_params;
132 int rem = get_optimize();
134 assert(n_tail_calls);
136 /* we add nwe nodes, so the outs are inconsistant */
137 set_irg_outs_inconsistent(irg);
139 /* we add new blocks and change the control flow */
140 set_irg_dom_inconsistent(irg);
142 /* we add a new loop */
143 set_irg_loopinfo_inconsistent(irg);
147 /* collect needed data */
152 data.proj_data = NULL;
153 irg_walk_graph(irg, NULL, collect_data, &data);
155 /* check number of arguments */
156 call = get_irn_link(end_block);
157 n_params = get_Call_n_params(call);
159 assert(data.proj_X && "Could not find initial exec from Start");
160 assert(data.block && "Could not find first block");
161 assert(data.proj_m && "Could not find ProjM(Start)");
162 assert((data.proj_data || n_params == 0) && "Could not find Proj(ProjT(Start)) of non-void function");
164 current_ir_graph = irg;
166 /* allocate in's for phi and block construction */
167 NEW_ARR_A(ir_node *, in, n_tail_calls + 1);
171 /* turn Return's into Jmp's */
172 for (i = 1, p = rets; p; p = get_irn_link(p)) {
175 set_cur_block(get_nodes_block(p));
178 exchange(p, new_Bad());
181 add_End_keepalive(get_irg_end(irg), jmp);
184 /* create a new block at start */
185 block = new_Block(n_tail_calls + 1, in);
188 /* the old first block is now the second one */
189 set_Block_cfgpred(data.block, data.blk_idx, jmp);
191 /* allocate phi's, position 0 contains the memory phi */
192 NEW_ARR_A(ir_node *, phis, n_params + 1);
194 /* build the memory phi */
196 in[i] = new_rd_Proj(NULL, irg, irg->start_block, irg->start, mode_M, pn_Start_M);
199 for (calls = call; calls; calls = get_irn_link(calls)) {
200 in[i] = get_Call_mem(calls);
203 assert(i == n_tail_calls + 1);
205 phis[0] = new_rd_Phi(NULL, irg, block, n_tail_calls + 1, in, mode_M);
207 /* build the data phi's */
211 NEW_ARR_A(ir_node **, call_params, n_params);
213 /* collect all parameters */
214 for (i = 0, calls = call; calls; calls = get_irn_link(calls)) {
215 call_params[i] = get_Call_param_arr(calls);
219 /* build new projs and Phi's */
220 for (i = 0; i < n_params; ++i) {
221 ir_mode *mode = get_irn_mode(call_params[0][i]);
223 in[0] = new_rd_Proj(NULL, irg, block, irg->args, mode, i);
224 for (j = 0; j < n_tail_calls; ++j)
225 in[j + 1] = call_params[j][i];
227 phis[i + 1] = new_rd_Phi(NULL, irg, block, n_tail_calls + 1, in, mode);
232 * ok, we are here, so we have build and collected all needed Phi's
233 * now exchange all Projs into links to Phi
235 for (p = data.proj_m; p; p = get_irn_link(p)) {
236 exchange(p, phis[0]);
238 for (p = data.proj_data; p; p = get_irn_link(p)) {
239 long proj = get_Proj_proj(p);
241 assert(0 <= proj && proj < n_params);
242 exchange(p, phis[proj + 1]);
245 /* tail recursion was done, all info is invalid */
246 set_irg_dom_inconsistent(irg);
247 set_irg_outs_inconsistent(irg);
248 set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent);
249 set_trouts_inconsistent();
250 set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
252 current_ir_graph = rem_irg;
257 * convert simple tail-calls into loops
259 int opt_tail_rec_irg(ir_graph *irg)
261 ir_node *end_block = irg->end_block;
263 int i, n_tail_calls = 0;
264 ir_node *rets = NULL;
266 if (! get_opt_tail_recursion() || ! get_opt_optimize())
269 set_irn_link(end_block, NULL);
271 n_preds = get_Block_n_cfgpreds(end_block);
272 for (i = 0; i < n_preds; ++i) {
273 ir_node *ret = get_Block_cfgpred(end_block, i);
274 ir_node *call, *call_ptr;
279 /* search all returns of a block */
280 if (get_irn_op(ret) != op_Return)
283 /* check, if it's a Return self() */
284 call = skip_Proj(get_Return_mem(ret));
285 if (get_irn_op(call) != op_Call)
288 /* check if it's a recursive call */
289 call_ptr = get_Call_ptr(call);
291 if (get_irn_op(call_ptr) != op_SymConst)
294 if (get_SymConst_kind(call_ptr) != symconst_addr_ent)
297 ent = get_SymConst_entity(call_ptr);
298 if (!ent || get_entity_irg(ent) != irg)
301 /* ok, mem is routed to a recursive call, check return args */
302 n_ress = get_Return_n_ress(ret);
303 ress = get_Return_res_arr(ret);
305 for (j = 0; j < n_ress; ++j) {
306 ir_node *irn = skip_Proj(skip_Proj(ress[j]));
309 /* not routed to a call */
316 /* here, we have found a call */
317 set_irn_link(call, get_irn_link(end_block));
318 set_irn_link(end_block, call);
321 /* link all returns, we will need this */
322 set_irn_link(ret, rets);
326 /* now, end_block->link contains the list of all tail calls */
330 if (get_opt_tail_recursion_verbose() && get_firm_verbosity() > 1)
331 printf(" Performing tail recursion for graph %s and %d Calls\n",
332 get_entity_ld_name(get_irg_entity(irg)), n_tail_calls);
334 do_opt_tail_rec(irg, rets, n_tail_calls);
340 * optimize tail recursion away
342 void opt_tail_recursion(void)
345 int n_opt_applications = 0;
347 if (! get_opt_tail_recursion() || ! get_opt_optimize())
350 for (i = 0; i < get_irp_n_irgs(); i++) {
351 current_ir_graph = get_irp_irg(i);
353 if (opt_tail_rec_irg(current_ir_graph))
354 ++n_opt_applications;
357 if (get_opt_tail_recursion_verbose())
358 printf("Performed tail recursion for %d of %d graphs\n", n_opt_applications, get_irp_n_irgs());