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"
34 #define set_bit(n) (returns[(n) >> 3] |= 1 << ((n) & 7))
35 #define get_bit(n) (returns[(n) >> 3] & (1 << ((n) & 7)))
38 #define IMAX(a, b) ((a) > (b) ? (a) : (b))
41 * Normalize the Returns of a graph by creating a new End block
42 * with One Return(Phi).
43 * This is the preferred input for the if-conversion.
45 * In pseudocode, it means:
60 void normalize_one_return(ir_graph *irg) {
61 ir_node *endbl = get_irg_end_block(irg);
62 int i, j, k, n, last_idx, n_rets, n_ret_vals = -1;
63 unsigned char *returns;
64 ir_node **in, **retvals, **endbl_in;
67 dbg_info *combined_dbgi = NULL;
69 /* look, if we have more than one return */
70 n = get_Block_n_cfgpreds(endbl);
72 /* The end block has no predecessors, we have an endless
73 loop. In that case, no returns exists. */
77 returns = ALLOCANZ(unsigned char, (n + 7) >> 3);
79 for (n_rets = i = 0; i < n; ++i) {
80 ir_node *node = get_Block_cfgpred(endbl, i);
82 if (is_Return(node)) {
83 dbg_info *dbgi = get_irn_dbg_info(node);
85 if (dbgi != NULL && dbgi != combined_dbgi) {
99 n_ret_vals = get_irn_arity(node);
103 /* there should be at least one Return node in Firm */
107 in = ALLOCAN(ir_node*, IMAX(n_rets, n_ret_vals));
108 retvals = ALLOCAN(ir_node*, n_rets * n_ret_vals);
109 endbl_in = ALLOCAN(ir_node*, n);
112 for (j = i = 0; i < n; ++i) {
113 ir_node *ret = get_Block_cfgpred(endbl, i);
116 ir_node *block = get_nodes_block(ret);
118 /* create a new Jmp for every Ret and place the in in */
119 in[j] = new_r_Jmp(irg, block);
121 /* save the return values and shuffle them */
122 for (k = 0; k < n_ret_vals; ++k)
123 retvals[j + k*n_rets] = get_irn_n(ret, k);
127 endbl_in[last_idx++] = ret;
130 /* ok, create a new block with all created in's */
131 block = new_r_Block(irg, n_rets, in);
133 /* now create the Phi nodes */
134 for (j = i = 0; i < n_ret_vals; ++i, j += n_rets) {
137 /* the return values are already shuffled */
139 /* Beware: normally the Phi constructor automatically replaces a Phi(a,...a) into a
140 but NOT, if a is Unknown. Here, we known that this case can be optimize also,
142 first = retvals[j + 0];
143 for (k = 1; k < n_rets; ++k) {
144 if (retvals[j + k] != first) {
152 in[i] = new_r_Phi(irg, block, n_rets, &retvals[j], get_irn_mode(retvals[j]));
155 endbl_in[last_idx++] = new_rd_Return(combined_dbgi, irg, block, in[0], n_ret_vals-1, &in[1]);
157 set_irn_in(endbl, last_idx, endbl_in);
159 /* invalidate analysis information:
160 * a new Block was added, so dominator, outs and loop are inconsistent,
161 * trouts and callee-state should be still valid
163 set_irg_doms_inconsistent(irg);
164 set_irg_outs_inconsistent(irg);
165 set_irg_extblk_inconsistent(irg);
166 set_irg_loopinfo_inconsistent(irg);
170 * Check, whether a Return can be moved on block upwards.
172 * In a block with a Return, all live nodes must be linked
173 * with the Return, otherwise they are dead (because the Return leaves
174 * the graph, so no more users of the other nodes can exists.
176 * We can move a Return, if it's predecessors are Phi nodes or
177 * comes from another block. In the later case, it is always possible
178 * to move the Return one block up, because the predecessor block must
179 * dominate the Return block (SSA) and then it dominates the predecessor
180 * block of the Return block as well.
182 * All predecessors of the Return block must be Jmp's of course, or we
183 * cannot move it up, so we add blocks if needed.
185 static int can_move_ret(ir_node *ret) {
186 ir_node *retbl = get_nodes_block(ret);
187 int i, n = get_irn_arity(ret);
189 for (i = 0; i < n; ++i) {
190 ir_node *pred = get_irn_n(ret, i);
192 if (! is_Phi(pred) && retbl == get_nodes_block(pred)) {
193 /* first condition failed, found a non-Phi predecessor
194 * then is in the Return block */
199 /* check, that predecessors are Jmps */
200 n = get_Block_n_cfgpreds(retbl);
203 for (i = 0; i < n; ++i) {
204 ir_node *pred = get_Block_cfgpred(retbl, i);
206 pred = skip_Tuple(pred);
207 if (! is_Jmp(pred) && !is_Bad(pred)) {
208 /* simply place a new block here */
209 ir_graph *irg = get_irn_irg(retbl);
210 ir_node *block = new_r_Block(irg, 1, &pred);
211 ir_node *jmp = new_r_Jmp(irg, block);
212 set_Block_cfgpred(retbl, i, jmp);
219 * Normalize the Returns of a graph by moving
220 * the Returns upwards as much as possible.
221 * This might be preferred for code generation.
223 * In pseudocode, it means:
231 * is transformed into
238 void normalize_n_returns(ir_graph *irg) {
239 int i, j, n, n_rets, n_finals, n_ret_vals;
240 ir_node *list = NULL;
241 ir_node *final = NULL;
243 ir_node *endbl = get_irg_end_block(irg);
247 * First, link all returns:
248 * These must be predecessors of the endblock.
249 * Place Returns that can be moved on list, all others
252 n = get_Block_n_cfgpreds(endbl);
253 for (n_finals = n_rets = i = 0; i < n; ++i) {
254 ir_node *ret = get_Block_cfgpred(endbl, i);
256 if (is_Return(ret) && can_move_ret(ret)) {
258 * Ok, all conditions met, we can move this Return, put it
261 set_irn_link(ret, list);
265 /* Put all nodes that are not changed on the final list. */
266 set_irn_link(ret, final);
276 * Now move the Returns upwards. We move always one block up (and create n
277 * new Returns), than we check if a newly created Return can be moved even further.
278 * If yes, we simply add it to our work list, else to the final list.
280 end = get_irg_end(irg);
281 n_ret_vals = get_irn_arity(list);
282 in = ALLOCAN(ir_node*, n_ret_vals);
285 ir_node *block = get_nodes_block(ret);
286 dbg_info *dbgi = get_irn_dbg_info(ret);
289 list = get_irn_link(ret);
292 n = get_Block_n_cfgpreds(block);
293 for (i = 0; i < n; ++i) {
294 ir_node *jmp = get_Block_cfgpred(block, i);
295 ir_node *new_bl, *new_ret;
301 new_bl = get_nodes_block(jmp);
303 /* create the in-array for the new Return */
304 for (j = 0; j < n_ret_vals; ++j) {
305 ir_node *pred = get_irn_n(ret, j);
307 in[j] = (is_Phi(pred) && get_nodes_block(pred) == block) ? get_Phi_pred(pred, i) : pred;
310 new_ret = new_rd_Return(dbgi, irg, new_bl, in[0], n_ret_vals - 1, &in[1]);
312 if (! is_Bad(new_ret)) {
314 * The newly created node might be bad, if we
315 * create it in a block with only Bad predecessors.
316 * In that case ignore this block.
318 * We could even kill the jmp then ...
320 if (can_move_ret(new_ret)) {
321 set_irn_link(new_ret, list);
325 set_irn_link(new_ret, final);
331 /* remove the Jmp, we have placed a Return here */
332 exchange(jmp, new_r_Bad(irg));
336 * if the memory of the old Return is a PhiM, remove it
337 * from the keep-alives, or it will keep the block which
338 * will crash the dominator algorithm.
340 phiM = get_Return_mem(ret);
342 n = get_End_n_keepalives(end);
343 for (i = 0; i < n; ++i) {
344 if (get_End_keepalive(end, i) == phiM) {
345 set_End_keepalive(end, i, new_r_Bad(irg));
353 * Last step: Create a new endblock, with all nodes on the final
354 * list as predecessors.
356 in = ALLOCAN(ir_node*, n_finals);
358 for (i = 0; final; ++i, final = get_irn_link(final))
361 exchange(endbl, new_r_Block(irg, n_finals, in));
363 /* the end block is not automatically skipped, so do it here */
364 set_irg_end_block(irg, skip_Id(get_irg_end_block(irg)));
366 /* Invalidate analysis information:
367 * Blocks become dead and new Returns were deleted, so dominator, outs and loop are inconsistent,
368 * trouts and callee-state should be still valid
370 set_irg_doms_inconsistent(irg);
371 set_irg_extblk_inconsistent(irg); /* may not be needed */
372 set_irg_outs_inconsistent(irg);
373 set_irg_loopinfo_inconsistent(current_ir_graph);