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)
63 ir_node *endbl = get_irg_end_block(irg);
64 int i, j, k, n, last_idx, n_rets, n_ret_vals = -1;
65 unsigned char *returns;
66 ir_node **in, **retvals, **endbl_in;
69 dbg_info *combined_dbgi = NULL;
71 /* look, if we have more than one return */
72 n = get_Block_n_cfgpreds(endbl);
74 /* The end block has no predecessors, we have an endless
75 loop. In that case, no returns exists. */
79 returns = ALLOCANZ(unsigned char, (n + 7) >> 3);
81 for (n_rets = i = 0; i < n; ++i) {
82 ir_node *node = get_Block_cfgpred(endbl, i);
84 if (is_Return(node)) {
85 dbg_info *dbgi = get_irn_dbg_info(node);
87 if (dbgi != NULL && dbgi != combined_dbgi) {
101 n_ret_vals = get_irn_arity(node);
105 /* there should be at least one Return node in Firm */
109 in = ALLOCAN(ir_node*, IMAX(n_rets, n_ret_vals));
110 retvals = ALLOCAN(ir_node*, n_rets * n_ret_vals);
111 endbl_in = ALLOCAN(ir_node*, n);
114 for (j = i = 0; i < n; ++i) {
115 ir_node *ret = get_Block_cfgpred(endbl, i);
118 ir_node *block = get_nodes_block(ret);
120 /* create a new Jmp for every Ret and place the in in */
121 in[j] = new_r_Jmp(block);
123 /* save the return values and shuffle them */
124 for (k = 0; k < n_ret_vals; ++k)
125 retvals[j + k*n_rets] = get_irn_n(ret, k);
129 endbl_in[last_idx++] = ret;
132 /* ok, create a new block with all created in's */
133 block = new_r_Block(irg, n_rets, in);
135 /* now create the Phi nodes */
136 for (j = i = 0; i < n_ret_vals; ++i, j += n_rets) {
139 /* the return values are already shuffled */
141 /* Beware: normally the Phi constructor automatically replaces a Phi(a,...a) into a
142 but NOT, if a is Unknown. Here, we known that this case can be optimize also,
144 first = retvals[j + 0];
145 for (k = 1; k < n_rets; ++k) {
146 if (retvals[j + k] != first) {
154 in[i] = new_r_Phi(block, n_rets, &retvals[j], get_irn_mode(retvals[j]));
157 endbl_in[last_idx++] = new_rd_Return(combined_dbgi, block, in[0], n_ret_vals-1, &in[1]);
159 set_irn_in(endbl, last_idx, endbl_in);
161 /* invalidate analysis information:
162 * a new Block was added, so dominator, outs and loop are inconsistent,
163 * trouts and callee-state should be still valid
165 set_irg_doms_inconsistent(irg);
166 set_irg_outs_inconsistent(irg);
167 set_irg_extblk_inconsistent(irg);
168 set_irg_loopinfo_inconsistent(irg);
171 /* Create a graph pass. */
172 ir_graph_pass_t *normalize_one_return_pass(const char *name)
174 return def_graph_pass(name ? name : "one_ret", normalize_one_return);
178 * Check, whether a Return can be moved on block upwards.
180 * In a block with a Return, all live nodes must be linked
181 * with the Return, otherwise they are dead (because the Return leaves
182 * the graph, so no more users of the other nodes can exists.
184 * We can move a Return, if its predecessors are Phi nodes or
185 * comes from another block. In the later case, it is always possible
186 * to move the Return one block up, because the predecessor block must
187 * dominate the Return block (SSA) and then it dominates the predecessor
188 * block of the Return block as well.
190 * All predecessors of the Return block must be Jmp's of course, or we
191 * cannot move it up, so we add blocks if needed.
193 static int can_move_ret(ir_node *ret)
195 ir_node *retbl = get_nodes_block(ret);
196 int i, n = get_irn_arity(ret);
198 for (i = 0; i < n; ++i) {
199 ir_node *pred = get_irn_n(ret, i);
201 if (! is_Phi(pred) && retbl == get_nodes_block(pred)) {
202 /* first condition failed, found a non-Phi predecessor
203 * then is in the Return block */
208 /* check, that predecessors are Jmps */
209 n = get_Block_n_cfgpreds(retbl);
210 /* we cannot move above a labeled block, as this might kill the block */
211 if (n <= 1 || has_Block_entity(retbl))
213 for (i = 0; i < n; ++i) {
214 ir_node *pred = get_Block_cfgpred(retbl, i);
216 pred = skip_Tuple(pred);
217 if (! is_Jmp(pred) && !is_Bad(pred)) {
218 /* simply place a new block here */
219 ir_graph *irg = get_irn_irg(retbl);
220 ir_node *block = new_r_Block(irg, 1, &pred);
221 ir_node *jmp = new_r_Jmp(block);
222 set_Block_cfgpred(retbl, i, jmp);
229 * Normalize the Returns of a graph by moving
230 * the Returns upwards as much as possible.
231 * This might be preferred for code generation.
233 * In pseudocode, it means:
241 * is transformed into
248 void normalize_n_returns(ir_graph *irg)
250 int i, j, n, n_rets, n_finals, n_ret_vals;
251 ir_node *list = NULL;
252 ir_node *final = NULL;
254 ir_node *endbl = get_irg_end_block(irg);
258 * First, link all returns:
259 * These must be predecessors of the endblock.
260 * Place Returns that can be moved on list, all others
263 n = get_Block_n_cfgpreds(endbl);
264 for (n_finals = n_rets = i = 0; i < n; ++i) {
265 ir_node *ret = get_Block_cfgpred(endbl, i);
267 if (is_Return(ret) && can_move_ret(ret)) {
269 * Ok, all conditions met, we can move this Return, put it
272 set_irn_link(ret, list);
276 /* Put all nodes that are not changed on the final list. */
277 set_irn_link(ret, final);
287 * Now move the Returns upwards. We move always one block up (and create n
288 * new Returns), than we check if a newly created Return can be moved even further.
289 * If yes, we simply add it to our work list, else to the final list.
291 end = get_irg_end(irg);
292 n_ret_vals = get_irn_arity(list);
293 in = ALLOCAN(ir_node*, n_ret_vals);
296 ir_node *block = get_nodes_block(ret);
297 dbg_info *dbgi = get_irn_dbg_info(ret);
300 list = (ir_node*)get_irn_link(ret);
303 n = get_Block_n_cfgpreds(block);
304 for (i = 0; i < n; ++i) {
305 ir_node *jmp = get_Block_cfgpred(block, i);
306 ir_node *new_bl, *new_ret;
312 new_bl = get_nodes_block(jmp);
314 /* create the in-array for the new Return */
315 for (j = 0; j < n_ret_vals; ++j) {
316 ir_node *pred = get_irn_n(ret, j);
318 in[j] = (is_Phi(pred) && get_nodes_block(pred) == block) ? get_Phi_pred(pred, i) : pred;
321 new_ret = new_rd_Return(dbgi, new_bl, in[0], n_ret_vals - 1, &in[1]);
323 if (! is_Bad(new_ret)) {
325 * The newly created node might be bad, if we
326 * create it in a block with only Bad predecessors.
327 * In that case ignore this block.
329 * We could even kill the jmp then ...
331 if (can_move_ret(new_ret)) {
332 set_irn_link(new_ret, list);
336 set_irn_link(new_ret, final);
342 /* remove the Jmp, we have placed a Return here */
343 exchange(jmp, new_r_Bad(irg));
347 * if the memory of the old Return is a PhiM, remove it
348 * from the keep-alives, or it will keep the block which
349 * will crash the dominator algorithm.
351 phiM = get_Return_mem(ret);
353 n = get_End_n_keepalives(end);
354 for (i = 0; i < n; ++i) {
355 if (get_End_keepalive(end, i) == phiM) {
356 set_End_keepalive(end, i, new_r_Bad(irg));
364 * Last step: Create a new endblock, with all nodes on the final
365 * list as predecessors.
367 in = ALLOCAN(ir_node*, n_finals);
369 for (i = 0; final != NULL; ++i, final = (ir_node*)get_irn_link(final))
372 exchange(endbl, new_r_Block(irg, n_finals, in));
374 /* the end block is not automatically skipped, so do it here */
375 set_irg_end_block(irg, skip_Id(get_irg_end_block(irg)));
377 /* Invalidate analysis information:
378 * Blocks become dead and new Returns were deleted, so dominator, outs and loop are inconsistent,
379 * trouts and callee-state should be still valid
381 set_irg_doms_inconsistent(irg);
382 set_irg_extblk_inconsistent(irg); /* may not be needed */
383 set_irg_outs_inconsistent(irg);
384 set_irg_loopinfo_inconsistent(irg);
387 /* Create a graph pass. */
388 ir_graph_pass_t *normalize_n_returns_pass(const char *name)
390 return def_graph_pass(name ? name : "n_rets", normalize_n_returns);