2 * Copyright (C) 1995-2008 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 Normalize returns.
23 * @author Michael Beck
28 #include "iroptimize.h"
29 #include "irgraph_t.h"
35 #define set_bit(n) (returns[(n) >> 3] |= 1 << ((n) & 7))
36 #define get_bit(n) (returns[(n) >> 3] & (1 << ((n) & 7)))
39 #define IMAX(a, b) ((a) > (b) ? (a) : (b))
42 * Normalize the Returns of a graph by creating a new End block
43 * with One Return(Phi).
44 * This is the preferred input for the if-conversion.
46 * In pseudocode, it means:
61 void normalize_one_return(ir_graph *irg) {
62 ir_node *endbl = get_irg_end_block(irg);
63 int i, j, k, n, last_idx, n_rets, n_ret_vals = -1;
64 unsigned char *returns;
65 ir_node **in, **retvals, **endbl_in;
68 dbg_info *combined_dbgi = NULL;
70 /* look, if we have more than one return */
71 n = get_Block_n_cfgpreds(endbl);
73 /* The end block has no predecessors, we have an endless
74 loop. In that case, no returns exists. */
78 returns = ALLOCANZ(unsigned char, (n + 7) >> 3);
80 for (n_rets = i = 0; i < n; ++i) {
81 ir_node *node = get_Block_cfgpred(endbl, i);
83 if (is_Return(node)) {
84 dbg_info *dbgi = get_irn_dbg_info(node);
86 if (dbgi != NULL && dbgi != combined_dbgi) {
100 n_ret_vals = get_irn_arity(node);
104 /* there should be at least one Return node in Firm */
108 in = ALLOCAN(ir_node*, IMAX(n_rets, n_ret_vals));
109 retvals = ALLOCAN(ir_node*, n_rets * n_ret_vals);
110 endbl_in = ALLOCAN(ir_node*, n);
113 for (j = i = 0; i < n; ++i) {
114 ir_node *ret = get_Block_cfgpred(endbl, i);
117 ir_node *block = get_nodes_block(ret);
119 /* create a new Jmp for every Ret and place the in in */
120 in[j] = new_r_Jmp(block);
122 /* save the return values and shuffle them */
123 for (k = 0; k < n_ret_vals; ++k)
124 retvals[j + k*n_rets] = get_irn_n(ret, k);
128 endbl_in[last_idx++] = ret;
131 /* ok, create a new block with all created in's */
132 block = new_r_Block(irg, n_rets, in);
134 /* now create the Phi nodes */
135 for (j = i = 0; i < n_ret_vals; ++i, j += n_rets) {
138 /* the return values are already shuffled */
140 /* Beware: normally the Phi constructor automatically replaces a Phi(a,...a) into a
141 but NOT, if a is Unknown. Here, we known that this case can be optimize also,
143 first = retvals[j + 0];
144 for (k = 1; k < n_rets; ++k) {
145 if (retvals[j + k] != first) {
153 in[i] = new_r_Phi(block, n_rets, &retvals[j], get_irn_mode(retvals[j]));
156 endbl_in[last_idx++] = new_rd_Return(combined_dbgi, block, in[0], n_ret_vals-1, &in[1]);
158 set_irn_in(endbl, last_idx, endbl_in);
160 /* invalidate analysis information:
161 * a new Block was added, so dominator, outs and loop are inconsistent,
162 * trouts and callee-state should be still valid
164 set_irg_doms_inconsistent(irg);
165 set_irg_outs_inconsistent(irg);
166 set_irg_extblk_inconsistent(irg);
167 set_irg_loopinfo_inconsistent(irg);
170 /* Create a graph pass. */
171 ir_graph_pass_t *normalize_one_return_pass(const char *name)
173 return def_graph_pass(name ? name : "one_ret", normalize_one_return);
177 * Check, whether a Return can be moved on block upwards.
179 * In a block with a Return, all live nodes must be linked
180 * with the Return, otherwise they are dead (because the Return leaves
181 * the graph, so no more users of the other nodes can exists.
183 * We can move a Return, if it's predecessors are Phi nodes or
184 * comes from another block. In the later case, it is always possible
185 * to move the Return one block up, because the predecessor block must
186 * dominate the Return block (SSA) and then it dominates the predecessor
187 * block of the Return block as well.
189 * All predecessors of the Return block must be Jmp's of course, or we
190 * cannot move it up, so we add blocks if needed.
192 static int can_move_ret(ir_node *ret) {
193 ir_node *retbl = get_nodes_block(ret);
194 int i, n = get_irn_arity(ret);
196 for (i = 0; i < n; ++i) {
197 ir_node *pred = get_irn_n(ret, i);
199 if (! is_Phi(pred) && retbl == get_nodes_block(pred)) {
200 /* first condition failed, found a non-Phi predecessor
201 * then is in the Return block */
206 /* check, that predecessors are Jmps */
207 n = get_Block_n_cfgpreds(retbl);
208 /* we cannot move above a labeled block, as this might kill the block */
209 if (n <= 1 || has_Block_entity(retbl))
211 for (i = 0; i < n; ++i) {
212 ir_node *pred = get_Block_cfgpred(retbl, i);
214 pred = skip_Tuple(pred);
215 if (! is_Jmp(pred) && !is_Bad(pred)) {
216 /* simply place a new block here */
217 ir_graph *irg = get_irn_irg(retbl);
218 ir_node *block = new_r_Block(irg, 1, &pred);
219 ir_node *jmp = new_r_Jmp(block);
220 set_Block_cfgpred(retbl, i, jmp);
227 * Normalize the Returns of a graph by moving
228 * the Returns upwards as much as possible.
229 * This might be preferred for code generation.
231 * In pseudocode, it means:
239 * is transformed into
246 void normalize_n_returns(ir_graph *irg) {
247 int i, j, n, n_rets, n_finals, n_ret_vals;
248 ir_node *list = NULL;
249 ir_node *final = NULL;
251 ir_node *endbl = get_irg_end_block(irg);
255 * First, link all returns:
256 * These must be predecessors of the endblock.
257 * Place Returns that can be moved on list, all others
260 n = get_Block_n_cfgpreds(endbl);
261 for (n_finals = n_rets = i = 0; i < n; ++i) {
262 ir_node *ret = get_Block_cfgpred(endbl, i);
264 if (is_Return(ret) && can_move_ret(ret)) {
266 * Ok, all conditions met, we can move this Return, put it
269 set_irn_link(ret, list);
273 /* Put all nodes that are not changed on the final list. */
274 set_irn_link(ret, final);
284 * Now move the Returns upwards. We move always one block up (and create n
285 * new Returns), than we check if a newly created Return can be moved even further.
286 * If yes, we simply add it to our work list, else to the final list.
288 end = get_irg_end(irg);
289 n_ret_vals = get_irn_arity(list);
290 in = ALLOCAN(ir_node*, n_ret_vals);
293 ir_node *block = get_nodes_block(ret);
294 dbg_info *dbgi = get_irn_dbg_info(ret);
297 list = get_irn_link(ret);
300 n = get_Block_n_cfgpreds(block);
301 for (i = 0; i < n; ++i) {
302 ir_node *jmp = get_Block_cfgpred(block, i);
303 ir_node *new_bl, *new_ret;
309 new_bl = get_nodes_block(jmp);
311 /* create the in-array for the new Return */
312 for (j = 0; j < n_ret_vals; ++j) {
313 ir_node *pred = get_irn_n(ret, j);
315 in[j] = (is_Phi(pred) && get_nodes_block(pred) == block) ? get_Phi_pred(pred, i) : pred;
318 new_ret = new_rd_Return(dbgi, new_bl, in[0], n_ret_vals - 1, &in[1]);
320 if (! is_Bad(new_ret)) {
322 * The newly created node might be bad, if we
323 * create it in a block with only Bad predecessors.
324 * In that case ignore this block.
326 * We could even kill the jmp then ...
328 if (can_move_ret(new_ret)) {
329 set_irn_link(new_ret, list);
333 set_irn_link(new_ret, final);
339 /* remove the Jmp, we have placed a Return here */
340 exchange(jmp, new_r_Bad(irg));
344 * if the memory of the old Return is a PhiM, remove it
345 * from the keep-alives, or it will keep the block which
346 * will crash the dominator algorithm.
348 phiM = get_Return_mem(ret);
350 n = get_End_n_keepalives(end);
351 for (i = 0; i < n; ++i) {
352 if (get_End_keepalive(end, i) == phiM) {
353 set_End_keepalive(end, i, new_r_Bad(irg));
361 * Last step: Create a new endblock, with all nodes on the final
362 * list as predecessors.
364 in = ALLOCAN(ir_node*, n_finals);
366 for (i = 0; final; ++i, final = get_irn_link(final))
369 exchange(endbl, new_r_Block(irg, n_finals, in));
371 /* the end block is not automatically skipped, so do it here */
372 set_irg_end_block(irg, skip_Id(get_irg_end_block(irg)));
374 /* Invalidate analysis information:
375 * Blocks become dead and new Returns were deleted, so dominator, outs and loop are inconsistent,
376 * trouts and callee-state should be still valid
378 set_irg_doms_inconsistent(irg);
379 set_irg_extblk_inconsistent(irg); /* may not be needed */
380 set_irg_outs_inconsistent(irg);
381 set_irg_loopinfo_inconsistent(current_ir_graph);
384 /* Create a graph pass. */
385 ir_graph_pass_t *normalize_n_returns_pass(const char *name)
387 return def_graph_pass(name ? name : "n_rets", normalize_n_returns);