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"
42 * the environment for collecting data
44 typedef struct _collect_t {
45 ir_node *proj_X; /**< initial exec proj */
46 ir_node *block; /**< old first block */
47 int blk_idx; /**< cfgpred index of the initial exec in block */
48 ir_node *proj_m; /**< linked list of memory from start proj's */
49 ir_node *proj_data; /**< linked list of all parameter access proj's */
53 * walker for collecting data, fills a collect_t environment
55 static void collect_data(ir_node *node, void *env)
57 collect_t *data = env;
61 switch (get_irn_opcode(node)) {
63 pred = get_Proj_pred(node);
65 op = get_irn_op(pred);
67 ir_node *start = get_Proj_pred(pred);
69 if (get_irn_op(start) == op_Start) {
70 if (get_Proj_proj(pred) == pn_Start_T_args) {
71 /* found Proj(ProjT(Start)) */
72 set_irn_link(node, data->proj_data);
73 data->proj_data = node;
77 else if (op == op_Start) {
78 switch (get_Proj_proj(node)) {
80 /* found ProjM(Start) */
81 set_irn_link(node, data->proj_m);
84 case pn_Start_X_initial_exec:
85 /* found ProjX(Start) */
94 int i, n_pred = get_Block_n_cfgpreds(node);
97 * the first block has the initial exec as cfg predecessor
99 if (node != current_ir_graph->start_block) {
100 for (i = 0; i < n_pred; ++i) {
101 if (get_Block_cfgpred(node, i) == data->proj_X) {
116 * do the graph reconstruction for tail-recursion elimination
118 * @param irg the graph that will reconstructed
119 * @param rets linked list of all rets
120 * @param n_tail_calls number of tail-recursion calls
122 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 entity *ent = get_irg_entity(irg);
134 type *method_tp = get_entity_type(ent);
136 assert(n_tail_calls);
138 /* we add new nodes, so the outs are inconsistant */
139 set_irg_outs_inconsistent(irg);
141 /* we add new blocks and change the control flow */
142 set_irg_dom_inconsistent(irg);
144 /* we add a new loop */
145 set_irg_loopinfo_inconsistent(irg);
147 /* calls are removed */
148 set_trouts_inconsistent();
150 /* we must build some new nodes WITHOUT CSE */
153 /* collect needed data */
158 data.proj_data = NULL;
159 irg_walk_graph(irg, NULL, collect_data, &data);
161 /* check number of arguments */
162 call = get_irn_link(end_block);
163 n_params = get_Call_n_params(call);
165 assert(data.proj_X && "Could not find initial exec from Start");
166 assert(data.block && "Could not find first block");
167 assert(data.proj_m && "Could not find ProjM(Start)");
168 assert((data.proj_data || n_params == 0) && "Could not find Proj(ProjT(Start)) of non-void function");
170 /* allocate in's for phi and block construction */
171 NEW_ARR_A(ir_node *, in, n_tail_calls + 1);
175 /* turn Return's into Jmp's */
176 for (i = 1, p = rets; p; p = get_irn_link(p)) {
178 ir_node *block = get_nodes_block(p);
180 jmp = new_r_Jmp(irg, block);
182 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);
205 for (calls = call; calls; calls = get_irn_link(calls)) {
206 in[i] = get_Call_mem(calls);
209 assert(i == n_tail_calls + 1);
211 phis[0] = new_r_Phi(irg, block, n_tail_calls + 1, in, mode_M);
213 /* 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 projs and Phi's */
226 for (i = 0; i < n_params; ++i) {
227 ir_mode *mode = get_type_mode(get_method_param_type(method_tp, i));
229 in[0] = new_r_Proj(irg, block, irg->args, mode, i);
230 for (j = 0; j < n_tail_calls; ++j)
231 in[j + 1] = call_params[j][i];
233 phis[i + 1] = new_r_Phi(irg, block, n_tail_calls + 1, in, mode);
238 * ok, we are here, so we have build and collected all needed Phi's
239 * now exchange all Projs into links to Phi
241 for (p = data.proj_m; p; p = get_irn_link(p)) {
242 exchange(p, phis[0]);
244 for (p = data.proj_data; p; p = get_irn_link(p)) {
245 long proj = get_Proj_proj(p);
247 assert(0 <= proj && proj < n_params);
248 exchange(p, phis[proj + 1]);
251 /* tail recursion was done, all info is invalid */
252 set_irg_dom_inconsistent(irg);
253 set_irg_outs_inconsistent(irg);
254 set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent);
255 set_trouts_inconsistent();
256 set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
262 * convert simple tail-calls into loops
264 int opt_tail_rec_irg(ir_graph *irg)
268 int i, n_tail_calls = 0;
269 ir_node *rets = NULL;
271 if (! get_opt_tail_recursion() || ! get_opt_optimize())
275 * This tail recursion optimization works best
276 * if the Returns are normalized.
278 normalize_n_returns(irg);
280 end_block = get_irg_end_block(irg);
281 set_irn_link(end_block, NULL);
283 n_preds = get_Block_n_cfgpreds(end_block);
284 for (i = 0; i < n_preds; ++i) {
285 ir_node *ret = get_Block_cfgpred(end_block, i);
286 ir_node *call, *call_ptr;
291 /* search all returns of a block */
292 if (get_irn_op(ret) != op_Return)
295 /* check, if it's a Return self() */
296 call = skip_Proj(get_Return_mem(ret));
297 if (get_irn_op(call) != op_Call)
300 /* check if it's a recursive call */
301 call_ptr = get_Call_ptr(call);
303 if (get_irn_op(call_ptr) != op_SymConst)
306 if (get_SymConst_kind(call_ptr) != symconst_addr_ent)
309 ent = get_SymConst_entity(call_ptr);
310 if (!ent || get_entity_irg(ent) != irg)
313 /* ok, mem is routed to a recursive call, check return args */
314 n_ress = get_Return_n_ress(ret);
315 ress = get_Return_res_arr(ret);
317 for (j = 0; j < n_ress; ++j) {
318 ir_node *irn = skip_Proj(skip_Proj(ress[j]));
321 /* not routed to a call */
328 /* here, we have found a call */
329 set_irn_link(call, get_irn_link(end_block));
330 set_irn_link(end_block, call);
333 /* link all returns, we will need this */
334 set_irn_link(ret, rets);
338 /* now, end_block->link contains the list of all tail calls */
342 if (get_opt_tail_recursion_verbose() && get_firm_verbosity() > 1)
343 printf(" Performing tail recursion for graph %s and %d Calls\n",
344 get_entity_ld_name(get_irg_entity(irg)), n_tail_calls);
346 hook_tail_rec(irg, n_tail_calls);
347 do_opt_tail_rec(irg, rets, n_tail_calls);
353 * optimize tail recursion away
355 void opt_tail_recursion(void)
358 int n_opt_applications = 0;
360 if (! get_opt_tail_recursion() || ! get_opt_optimize())
363 for (i = 0; i < get_irp_n_irgs(); i++) {
364 current_ir_graph = get_irp_irg(i);
366 if (opt_tail_rec_irg(current_ir_graph))
367 ++n_opt_applications;
370 if (get_opt_tail_recursion_verbose())
371 printf("Performed tail recursion for %d of %d graphs\n", n_opt_applications, get_irp_n_irgs());