2 * Copyright (C) 1995-2011 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
32 #include "iroptimize.h"
33 #include "scalar_replace.h"
40 #include "irgraph_t.h"
48 #include "opt_manage.h"
50 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
53 * the environment for collecting data
55 typedef struct collect_t {
56 ir_node *proj_X; /**< initial exec proj */
57 ir_node *block; /**< old first block */
58 int blk_idx; /**< cfgpred index of the initial exec in block */
59 ir_node *proj_m; /**< memory from start proj's */
60 ir_node *proj_data; /**< linked list of all parameter access proj's */
64 * walker for collecting data, fills a collect_t environment
66 static void collect_data(ir_node *node, void *env)
68 collect_t *data = (collect_t*)env;
72 switch (get_irn_opcode(node)) {
74 pred = get_Proj_pred(node);
76 opcode = get_irn_opcode(pred);
77 if (opcode == iro_Proj) {
78 ir_node *start = get_Proj_pred(pred);
80 if (is_Start(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 (opcode == iro_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(get_irn_irg(node))) {
101 for (i = 0; i < n_pred; ++i) {
102 if (get_Block_cfgpred(node, i) == data->proj_X) {
116 typedef enum tail_rec_variants {
117 TR_DIRECT, /**< direct return value, i.e. return func(). */
118 TR_ADD, /**< additive return value, i.e. return x +/- func() */
119 TR_MUL, /**< multiplicative return value, i.e. return x * func() or return -func() */
120 TR_BAD, /**< any other transformation */
121 TR_UNKNOWN /**< during construction */
124 typedef struct tr_env {
125 int n_tail_calls; /**< number of tail calls found */
126 int n_ress; /**< number of return values */
127 tail_rec_variants *variants; /**< return value variants */
128 ir_node *rets; /**< list of returns that can be transformed */
133 * do the graph reconstruction for tail-recursion elimination
135 * @param irg the graph that will reconstructed
136 * @param env tail recursion environment
138 static void do_opt_tail_rec(ir_graph *irg, tr_env *env)
140 ir_node *end_block = get_irg_end_block(irg);
141 ir_node *block, *jmp, *call, *calls;
144 ir_node ***call_params;
146 int i, j, n_params, n_locs;
148 int rem = get_optimize();
149 ir_entity *ent = get_irg_entity(irg);
150 ir_type *method_tp = get_entity_type(ent);
152 assert(env->n_tail_calls > 0);
154 /* we add new blocks and change the control flow */
155 clear_irg_state(irg, IR_GRAPH_STATE_CONSISTENT_DOMINANCE
156 | IR_GRAPH_STATE_VALID_EXTENDED_BLOCKS);
158 /* we must build some new nodes WITHOUT CSE */
161 /* collect needed data */
165 data.proj_m = get_irg_initial_mem(irg);
166 data.proj_data = NULL;
167 irg_walk_graph(irg, NULL, collect_data, &data);
169 /* check number of arguments */
170 call = (ir_node*)get_irn_link(end_block);
171 n_params = get_Call_n_params(call);
173 assert(data.proj_X && "Could not find initial exec from Start");
174 assert(data.block && "Could not find first block");
175 assert(data.proj_m && "Could not find initial memory");
176 assert((data.proj_data || n_params == 0) && "Could not find Proj(ProjT(Start)) of non-void function");
178 /* allocate in's for phi and block construction */
179 NEW_ARR_A(ir_node *, in, env->n_tail_calls + 1);
181 /* build a new header block for the loop we create */
183 in[i++] = data.proj_X;
185 /* turn Return's into Jmp's */
186 for (p = env->rets; p; p = n) {
187 ir_node *block = get_nodes_block(p);
189 n = (ir_node*)get_irn_link(p);
190 in[i++] = new_r_Jmp(block);
192 // exchange(p, new_r_Bad(irg));
194 /* we might generate an endless loop, so add
195 * the block to the keep-alive list */
196 add_End_keepalive(get_irg_end(irg), block);
198 assert(i == env->n_tail_calls + 1);
201 block = new_r_Block(irg, i, in);
202 jmp = new_r_Jmp(block);
204 /* the old first block is now the second one */
205 set_Block_cfgpred(data.block, data.blk_idx, jmp);
207 /* allocate phi's, position 0 contains the memory phi */
208 NEW_ARR_A(ir_node *, phis, n_params + 1);
210 /* build the memory phi */
212 in[i] = new_r_Proj(get_irg_start(irg), mode_M, pn_Start_M);
213 set_irg_initial_mem(irg, in[i]);
216 for (calls = call; calls != NULL; calls = (ir_node*)get_irn_link(calls)) {
217 in[i] = get_Call_mem(calls);
220 assert(i == env->n_tail_calls + 1);
222 phis[0] = new_r_Phi(block, env->n_tail_calls + 1, in, mode_M);
224 /* build the data Phi's */
229 NEW_ARR_A(ir_node **, call_params, env->n_tail_calls);
231 /* collect all parameters */
232 for (i = 0, calls = call; calls != NULL;
233 calls = (ir_node*)get_irn_link(calls)) {
234 call_params[i] = get_Call_param_arr(calls);
238 /* build new Proj's and Phi's */
239 args = get_irg_args(irg);
240 for (i = 0; i < n_params; ++i) {
241 ir_mode *mode = get_type_mode(get_method_param_type(method_tp, i));
243 in[0] = new_r_Proj(args, mode, i);
244 for (j = 0; j < env->n_tail_calls; ++j)
245 in[j + 1] = call_params[j][i];
247 phis[i + 1] = new_r_Phi(block, env->n_tail_calls + 1, in, mode);
252 * ok, we are here, so we have build and collected all needed Phi's
253 * now exchange all Projs into links to Phi
255 exchange(data.proj_m, 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);
260 n = (ir_node*)get_irn_link(p);
261 exchange(p, phis[proj + 1]);
264 /* tail recursion was done, all info is invalid */
265 clear_irg_state(irg, IR_GRAPH_STATE_CONSISTENT_DOMINANCE
266 | IR_GRAPH_STATE_CONSISTENT_LOOPINFO
267 | IR_GRAPH_STATE_VALID_EXTENDED_BLOCKS);
268 set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
272 /* check if we need new values */
274 for (i = 0; i < env->n_ress; ++i) {
275 if (env->variants[i] != TR_DIRECT) {
282 ir_node *start_block;
286 NEW_ARR_A(ir_node *, in, env->n_ress);
287 NEW_ARR_A(ir_mode *, modes, env->n_ress);
288 ssa_cons_start(irg, env->n_ress);
290 start_block = get_irg_start_block(irg);
291 set_r_cur_block(irg, start_block);
293 /* set the neutral elements for the iteration start */
294 for (i = 0; i < env->n_ress; ++i) {
295 ir_type *tp = get_method_res_type(method_tp, i);
296 ir_mode *mode = get_type_mode(tp);
299 if (env->variants[i] == TR_ADD) {
300 set_r_value(irg, i, new_r_Const(irg, get_mode_null(mode)));
301 } else if (env->variants[i] == TR_MUL) {
302 set_r_value(irg, i, new_r_Const(irg, get_mode_one(mode)));
305 mature_immBlock(start_block);
307 /* no: we can kill all returns */
308 for (p = env->rets; p; p = n) {
309 ir_node *block = get_nodes_block(p);
310 ir_node *call, *mem, *jmp, *tuple;
312 set_r_cur_block(irg, block);
313 n = (ir_node*)get_irn_link(p);
315 call = skip_Proj(get_Return_mem(p));
316 assert(is_Call(call));
318 mem = get_Call_mem(call);
320 /* create a new jump, free of CSE */
322 jmp = new_r_Jmp(block);
325 for (i = 0; i < env->n_ress; ++i) {
326 ir_mode *mode = modes[i];
327 if (env->variants[i] != TR_DIRECT) {
328 in[i] = get_r_value(irg, i, mode);
330 in[i] = new_r_Bad(irg, mode);
333 /* create a new tuple for the return values */
334 tuple = new_r_Tuple(block, env->n_ress, in);
336 turn_into_tuple(call, pn_Call_max+1);
337 set_Tuple_pred(call, pn_Call_M, mem);
338 set_Tuple_pred(call, pn_Call_X_regular, jmp);
339 set_Tuple_pred(call, pn_Call_X_except, new_r_Bad(irg, mode_X));
340 set_Tuple_pred(call, pn_Call_T_result, tuple);
342 for (i = 0; i < env->n_ress; ++i) {
343 ir_node *res = get_Return_res(p, i);
344 if (env->variants[i] != TR_DIRECT) {
345 set_r_value(irg, i, res);
349 exchange(p, new_r_Bad(irg, mode_X));
352 /* finally fix all other returns */
353 end_block = get_irg_end_block(irg);
354 for (i = get_Block_n_cfgpreds(end_block) - 1; i >= 0; --i) {
355 ir_node *ret = get_Block_cfgpred(end_block, i);
358 /* search all Returns of a block */
359 if (! is_Return(ret))
362 block = get_nodes_block(ret);
363 set_r_cur_block(irg, block);
364 for (j = 0; j < env->n_ress; ++j) {
365 ir_node *pred = get_Return_res(ret, j);
368 switch (env->variants[j]) {
373 n = get_r_value(irg, j, modes[j]);
374 n = new_r_Add(block, n, pred, modes[j]);
375 set_Return_res(ret, j, n);
379 n = get_r_value(irg, j, modes[j]);
380 n = new_r_Mul(block, n, pred, modes[j]);
381 set_Return_res(ret, j, n);
385 assert(!"unexpected tail recursion variant");
389 ssa_cons_finish(irg);
391 ir_node *bad = new_r_Bad(irg, mode_X);
393 /* no: we can kill all returns */
394 for (p = env->rets; p; p = n) {
395 n = (ir_node*)get_irn_link(p);
402 * Check the lifetime of locals in the given graph.
403 * Tail recursion can only be done, if we can prove that
404 * the lifetime of locals end with the recursive call.
405 * We do this by checking that no address of a local variable is
406 * stored or transmitted as an argument to a call.
408 * @return non-zero if it's ok to do tail recursion
410 static int check_lifetime_of_locals(ir_graph *irg)
414 ir_type *frame_tp = get_irg_frame_type(irg);
416 irg_frame = get_irg_frame(irg);
417 for (i = get_irn_n_outs(irg_frame) - 1; i >= 0; --i) {
418 ir_node *succ = get_irn_out(irg_frame, i);
421 /* Check if we have compound arguments.
422 For now, we cannot handle them, */
423 if (get_entity_owner(get_Sel_entity(succ)) != frame_tp)
426 if (is_address_taken(succ))
434 * Examine irn and detect the recursion variant.
436 static tail_rec_variants find_variant(ir_node *irn, ir_node *call)
439 tail_rec_variants va, vb, res;
441 if (skip_Proj(skip_Proj(irn)) == call) {
445 switch (get_irn_opcode(irn)) {
448 a = get_Add_left(irn);
449 if (get_nodes_block(a) != get_nodes_block(call)) {
450 /* we are outside, ignore */
453 va = find_variant(a, call);
457 b = get_Add_right(irn);
458 if (get_nodes_block(b) != get_nodes_block(call)) {
459 /* we are outside, ignore */
462 vb = find_variant(b, call);
469 else if (va == TR_UNKNOWN)
471 else if (vb == TR_UNKNOWN)
474 /* they are different but none is TR_UNKNOWN -> incompatible */
477 if (res == TR_DIRECT || res == TR_ADD)
483 /* try additive, but return value must be left */
484 a = get_Sub_left(irn);
485 if (get_nodes_block(a) != get_nodes_block(call)) {
486 /* we are outside, ignore */
489 va = find_variant(a, call);
493 b = get_Sub_right(irn);
494 if (get_nodes_block(b) != get_nodes_block(call)) {
495 /* we are outside, ignore */
498 vb = find_variant(b, call);
499 if (vb != TR_UNKNOWN)
503 if (res == TR_DIRECT || res == TR_ADD)
509 /* try multiplicative */
510 a = get_Mul_left(irn);
511 if (get_nodes_block(a) != get_nodes_block(call)) {
512 /* we are outside, ignore */
515 va = find_variant(a, call);
519 b = get_Mul_right(irn);
520 if (get_nodes_block(b) != get_nodes_block(call)) {
521 /* we are outside, ignore */
524 vb = find_variant(b, call);
531 else if (va == TR_UNKNOWN)
533 else if (vb == TR_UNKNOWN)
536 /* they are different but none is TR_UNKNOWN -> incompatible */
539 if (res == TR_DIRECT || res == TR_MUL)
545 /* try multiplicative */
546 a = get_Minus_op(irn);
547 res = find_variant(a, call);
548 if (res == TR_DIRECT)
550 if (res == TR_MUL || res == TR_UNKNOWN)
562 * convert simple tail-calls into loops
564 static ir_graph_state_t do_tailrec(ir_graph *irg)
568 int i, n_ress, n_tail_calls = 0;
569 ir_node *rets = NULL;
570 ir_type *mtd_type, *call_type;
574 FIRM_DBG_REGISTER(dbg, "firm.opt.tailrec");
576 if (! check_lifetime_of_locals(irg))
579 rem = current_ir_graph;
580 current_ir_graph = irg;
582 ent = get_irg_entity(irg);
583 mtd_type = get_entity_type(ent);
584 n_ress = get_method_n_ress(mtd_type);
590 NEW_ARR_A(tail_rec_variants, env.variants, n_ress);
592 for (i = 0; i < n_ress; ++i)
593 env.variants[i] = TR_DIRECT;
596 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
598 end_block = get_irg_end_block(irg);
599 set_irn_link(end_block, NULL);
601 for (i = get_Block_n_cfgpreds(end_block) - 1; i >= 0; --i) {
602 ir_node *ret = get_Block_cfgpred(end_block, i);
603 ir_node *call, *call_ptr;
607 /* search all Returns of a block */
608 if (! is_Return(ret))
611 /* check, if it's a Return self() */
612 call = skip_Proj(get_Return_mem(ret));
616 /* the call must be in the same block as the return */
617 if (get_nodes_block(call) != get_nodes_block(ret))
620 /* check if it's a recursive call */
621 call_ptr = get_Call_ptr(call);
623 if (! is_SymConst_addr_ent(call_ptr))
626 ent = get_SymConst_entity(call_ptr);
627 if (!ent || get_entity_irg(ent) != irg)
631 * Check, that the types match. At least in C
634 mtd_type = get_entity_type(ent);
635 call_type = get_Call_type(call);
637 if (mtd_type != call_type) {
639 * Hmm, the types did not match, bad.
640 * This can happen in C when no prototype is given
641 * or K&R style is used.
643 DB((dbg, LEVEL_3, " tail recursion fails because of call type mismatch: %+F != %+F\n", mtd_type, call_type));
647 /* ok, mem is routed to a recursive call, check return args */
648 ress = get_Return_res_arr(ret);
649 for (j = get_Return_n_ress(ret) - 1; j >= 0; --j) {
650 tail_rec_variants var = find_variant(ress[j], call);
653 /* cannot be transformed */
656 if (var == TR_DIRECT)
657 var = env.variants[j];
658 else if (env.variants[j] == TR_DIRECT)
659 env.variants[j] = var;
660 if (env.variants[j] != var) {
662 DB((dbg, LEVEL_3, " tail recursion fails for %d return value of %+F\n", j, ret));
669 /* here, we have found a call */
670 set_irn_link(call, get_irn_link(end_block));
671 set_irn_link(end_block, call);
674 /* link all returns, we will need this */
675 set_irn_link(ret, rets);
679 /* now, end_block->link contains the list of all tail calls */
680 if (n_tail_calls > 0) {
681 DB((dbg, LEVEL_2, " Performing tail recursion for graph %s and %d Calls\n",
682 get_entity_ld_name(get_irg_entity(irg)), n_tail_calls));
684 hook_tail_rec(irg, n_tail_calls);
686 env.n_tail_calls = n_tail_calls;
688 do_opt_tail_rec(irg, &env);
690 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
691 current_ir_graph = rem;
697 * This tail recursion optimization works best
698 * if the Returns are normalized.
700 static optdesc_t opt_tailrec = {
702 IR_GRAPH_STATE_MANY_RETURNS | IR_GRAPH_STATE_NO_BADS | IR_GRAPH_STATE_CONSISTENT_OUTS,
706 int opt_tail_rec_irg(ir_graph *irg) {
707 perform_irg_optimization(irg, &opt_tailrec);
708 return 1; /* conservatively report changes */
711 ir_graph_pass_t *opt_tail_rec_irg_pass(const char *name)
713 return def_graph_pass_ret(name ? name : "tailrec", opt_tail_rec_irg);
717 * optimize tail recursion away
719 void opt_tail_recursion(void)
722 size_t n_opt_applications = 0;
724 FIRM_DBG_REGISTER(dbg, "firm.opt.tailrec");
726 DB((dbg, LEVEL_1, "Performing tail recursion ...\n"));
727 for (i = 0, n = get_irp_n_irgs(); i < n; ++i) {
728 ir_graph *irg = get_irp_irg(i);
730 if (opt_tail_rec_irg(irg))
731 ++n_opt_applications;
734 DB((dbg, LEVEL_1, "Done for %zu of %zu graphs.\n",
735 n_opt_applications, get_irp_n_irgs()));
738 ir_prog_pass_t *opt_tail_recursion_pass(const char *name)
740 return def_prog_pass(name ? name : "tailrec", opt_tail_recursion);