2 * This file is part of libFirm.
3 * Copyright (C) 2012 University of Karlsruhe.
8 * @brief Normalize returns.
15 #include "iroptimize.h"
16 #include "irgraph_t.h"
22 #include "raw_bitset.h"
25 * Normalize the Returns of a graph by creating a new End block
26 * with One Return(Phi).
27 * This is the preferred input for the if-conversion.
29 * In pseudocode, it means:
44 void normalize_one_return(ir_graph *irg)
46 ir_node *endbl = get_irg_end_block(irg);
47 ir_entity *entity = get_irg_entity(irg);
48 ir_type *type = get_entity_type(entity);
49 int n_ret_vals = get_method_n_ress(type) + 1;
51 bool filter_dbgi = false;
52 dbg_info *combined_dbgi = NULL;
53 int i, j, k, n, last_idx;
54 ir_node **in, **retvals, **endbl_in;
57 /* look, if we have more than one return */
58 n = get_Block_n_cfgpreds(endbl);
60 /* The end block has no predecessors, we have an endless
61 loop. In that case, no returns exists. */
62 confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_ALL);
63 add_irg_properties(irg, IR_GRAPH_PROPERTY_ONE_RETURN);
67 unsigned *const returns = rbitset_alloca(n);
68 for (i = 0; i < n; ++i) {
69 ir_node *node = get_Block_cfgpred(endbl, i);
71 if (is_Return(node)) {
72 dbg_info *dbgi = get_irn_dbg_info(node);
74 if (dbgi != NULL && dbgi != combined_dbgi) {
84 rbitset_set(returns, i);
89 confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_ALL);
90 add_irg_properties(irg, IR_GRAPH_PROPERTY_ONE_RETURN);
94 in = ALLOCAN(ir_node*, MAX(n_rets, n_ret_vals));
95 retvals = ALLOCAN(ir_node*, n_rets * n_ret_vals);
96 endbl_in = ALLOCAN(ir_node*, n);
99 for (j = i = 0; i < n; ++i) {
100 ir_node *ret = get_Block_cfgpred(endbl, i);
102 if (rbitset_is_set(returns, i)) {
103 ir_node *block = get_nodes_block(ret);
105 /* create a new Jmp for every Ret and place the in in */
106 in[j] = new_r_Jmp(block);
108 /* save the return values and shuffle them */
109 for (k = 0; k < n_ret_vals; ++k)
110 retvals[j + k*n_rets] = get_irn_n(ret, k);
114 endbl_in[last_idx++] = ret;
118 /* ok, create a new block with all created in's */
119 block = new_r_Block(irg, n_rets, in);
121 /* now create the Phi nodes */
122 for (j = i = 0; i < n_ret_vals; ++i, j += n_rets) {
123 ir_mode *mode = get_irn_mode(retvals[j]);
124 in[i] = new_r_Phi(block, n_rets, &retvals[j], mode);
127 endbl_in[last_idx++] = new_rd_Return(combined_dbgi, block, in[0], n_ret_vals-1, &in[1]);
129 set_irn_in(endbl, last_idx, endbl_in);
131 /* invalidate analysis information:
132 * a new Block was added, so dominator, outs and loop are inconsistent,
133 * trouts and callee-state should be still valid */
134 confirm_irg_properties(irg,
135 IR_GRAPH_PROPERTY_NO_BADS
136 | IR_GRAPH_PROPERTY_NO_TUPLES
137 | IR_GRAPH_PROPERTY_NO_CRITICAL_EDGES
138 | IR_GRAPH_PROPERTY_NO_UNREACHABLE_CODE
139 | IR_GRAPH_PROPERTY_CONSISTENT_ENTITY_USAGE);
140 add_irg_properties(irg, IR_GRAPH_PROPERTY_ONE_RETURN);
143 /* Create a graph pass. */
144 ir_graph_pass_t *normalize_one_return_pass(const char *name)
146 return def_graph_pass(name ? name : "one_ret", normalize_one_return);
150 * Check, whether a Return can be moved on block upwards.
152 * In a block with a Return, all live nodes must be linked
153 * with the Return, otherwise they are dead (because the Return leaves
154 * the graph, so no more users of the other nodes can exists.
156 * We can move a Return, if its predecessors are Phi nodes or
157 * comes from another block. In the later case, it is always possible
158 * to move the Return one block up, because the predecessor block must
159 * dominate the Return block (SSA) and then it dominates the predecessor
160 * block of the Return block as well.
162 * All predecessors of the Return block must be Jmp's of course, or we
163 * cannot move it up, so we add blocks if needed.
165 static bool can_move_ret(ir_node *ret)
167 ir_node *retbl = get_nodes_block(ret);
168 int i, n = get_irn_arity(ret);
170 for (i = 0; i < n; ++i) {
171 ir_node *pred = get_irn_n(ret, i);
173 if (! is_Phi(pred) && retbl == get_nodes_block(pred)) {
174 /* first condition failed, found a non-Phi predecessor
175 * then is in the Return block */
180 /* check, that predecessors are Jmps */
181 n = get_Block_n_cfgpreds(retbl);
182 /* we cannot move above a labeled block, as this might kill the block */
183 if (n <= 1 || get_Block_entity(retbl) != NULL)
185 for (i = 0; i < n; ++i) {
186 ir_node *pred = get_Block_cfgpred(retbl, i);
188 pred = skip_Tuple(pred);
189 if (! is_Jmp(pred) && !is_Bad(pred)) {
190 /* simply place a new block here */
191 ir_graph *irg = get_irn_irg(retbl);
192 ir_node *block = new_r_Block(irg, 1, &pred);
193 ir_node *jmp = new_r_Jmp(block);
194 set_Block_cfgpred(retbl, i, jmp);
201 * Normalize the Returns of a graph by moving
202 * the Returns upwards as much as possible.
203 * This might be preferred for code generation.
205 * In pseudocode, it means:
213 * is transformed into
220 void normalize_n_returns(ir_graph *irg)
223 ir_node *list = NULL;
224 ir_node *final = NULL;
226 unsigned n_finals = 0;
227 ir_node *endbl = get_irg_end_block(irg);
233 * First, link all returns:
234 * These must be predecessors of the endblock.
235 * Place Returns that can be moved on list, all others
238 n = get_Block_n_cfgpreds(endbl);
239 for (i = 0; i < n; ++i) {
240 ir_node *ret = get_Block_cfgpred(endbl, i);
244 } else if (is_Return(ret) && can_move_ret(ret)) {
246 * Ok, all conditions met, we can move this Return, put it
249 set_irn_link(ret, list);
253 /* Put all nodes that are not changed on the final list. */
254 set_irn_link(ret, final);
261 confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_ALL);
262 add_irg_properties(irg, IR_GRAPH_PROPERTY_MANY_RETURNS);
267 * Now move the Returns upwards. We move always one block up (and create n
268 * new Returns), than we check if a newly created Return can be moved even
269 * further. If yes, we simply add it to our work list, else to the final
272 end = get_irg_end(irg);
273 n_ret_vals = get_irn_arity(list);
274 in = ALLOCAN(ir_node*, n_ret_vals);
275 while (list != NULL) {
277 ir_node *block = get_nodes_block(ret);
278 dbg_info *dbgi = get_irn_dbg_info(ret);
281 list = (ir_node*)get_irn_link(ret);
284 n = get_Block_n_cfgpreds(block);
285 for (i = 0; i < n; ++i) {
286 ir_node *jmp = get_Block_cfgpred(block, i);
287 ir_node *new_bl, *new_ret;
293 new_bl = get_nodes_block(jmp);
295 /* create the in-array for the new Return */
296 for (j = 0; j < n_ret_vals; ++j) {
297 ir_node *pred = get_irn_n(ret, j);
299 in[j] = (is_Phi(pred) && get_nodes_block(pred) == block) ? get_Phi_pred(pred, i) : pred;
302 new_ret = new_rd_Return(dbgi, new_bl, in[0], n_ret_vals-1, &in[1]);
304 if (! is_Bad(new_ret)) {
306 * The newly created node might be bad, if we
307 * create it in a block with only Bad predecessors.
308 * In that case ignore this block.
310 * We could even kill the jmp then ...
312 if (can_move_ret(new_ret)) {
313 set_irn_link(new_ret, list);
317 set_irn_link(new_ret, final);
323 /* remove the Jmp, we have placed a Return here */
324 exchange(jmp, new_r_Bad(irg, mode_X));
328 * if the memory of the old Return is a PhiM, remove it
329 * from the keep-alives, or it will keep the block which
330 * will crash the dominator algorithm.
332 phiM = get_Return_mem(ret);
334 n = get_End_n_keepalives(end);
335 for (i = 0; i < n; ++i) {
336 if (get_End_keepalive(end, i) == phiM) {
337 set_End_keepalive(end, i, new_r_Bad(irg, mode_M));
345 * Last step: Create a new endblock, with all nodes on the final list as
348 in = ALLOCAN(ir_node*, n_finals);
350 for (i = 0; final != NULL; ++i, final = (ir_node*)get_irn_link(final)) {
354 exchange(endbl, new_r_Block(irg, n_finals, in));
356 /* Invalidate analysis information:
357 * Blocks become dead and new Returns were deleted, so dominator, outs and
358 * loop are inconsistent, trouts and callee-state should be still valid */
359 confirm_irg_properties(irg,
360 IR_GRAPH_PROPERTY_NO_TUPLES
361 | IR_GRAPH_PROPERTY_NO_CRITICAL_EDGES
362 | IR_GRAPH_PROPERTY_CONSISTENT_ENTITY_USAGE);
363 add_irg_properties(irg, IR_GRAPH_PROPERTY_MANY_RETURNS);
366 /* Create a graph pass. */
367 ir_graph_pass_t *normalize_n_returns_pass(const char *name)
369 return def_graph_pass(name ? name : "n_rets", normalize_n_returns);