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;
68 /* look, if we have more than one return */
69 n = get_Block_n_cfgpreds(endbl);
71 /* The end block has no predecessors, we have an endless
72 loop. In that case, no returns exists. */
76 returns = alloca((n + 7) >> 3);
77 memset(returns, 0, (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)) {
88 n_ret_vals = get_irn_arity(node);
92 /* there should be at least one Return node in Firm */
96 in = alloca(sizeof(*in) * IMAX(n_rets, n_ret_vals));
97 retvals = alloca(sizeof(*retvals) * n_rets * n_ret_vals);
98 endbl_in = alloca(sizeof(*endbl_in) * n);
101 for (j = i = 0; i < n; ++i) {
102 ir_node *ret = get_Block_cfgpred(endbl, i);
105 ir_node *block = get_nodes_block(ret);
107 /* create a new Jmp for every Ret and place the in in */
108 in[j] = new_r_Jmp(irg, block);
110 /* save the return values and shuffle them */
111 for (k = 0; k < n_ret_vals; ++k)
112 retvals[j + k*n_rets] = get_irn_n(ret, k);
116 endbl_in[last_idx++] = ret;
119 /* ok, create a new block with all created in's */
120 block = new_r_Block(irg, n_rets, in);
122 /* now create the Phi nodes */
123 for (j = i = 0; i < n_ret_vals; ++i, j += n_rets) {
126 /* the return values are already shuffled */
128 /* Beware: normally the Phi constructor automatically replaces a Phi(a,...a) into a
129 but NOT, if a is Unknown. Here, we known that this case can be optimize also,
131 first = retvals[j + 0];
132 for (k = 1; k < n_rets; ++k) {
133 if (retvals[j + k] != first) {
141 in[i] = new_r_Phi(irg, block, n_rets, &retvals[j], get_irn_mode(retvals[j]));
144 endbl_in[last_idx++] = new_r_Return(irg, block, in[0], n_ret_vals-1, &in[1]);
146 set_irn_in(endbl, last_idx, endbl_in);
148 /* invalidate analysis information:
149 * a new Block was added, so dominator, outs and loop are inconsistent,
150 * trouts and callee-state should be still valid
152 set_irg_doms_inconsistent(irg);
153 set_irg_outs_inconsistent(irg);
154 set_irg_extblk_inconsistent(irg);
155 set_irg_loopinfo_inconsistent(irg);
159 * Check, whether a Return can be moved on block upwards.
161 * In a block with a Return, all live nodes must be linked
162 * with the Return, otherwise they are dead (because the Return leaves
163 * the graph, so no more users of the other nodes can exists.
165 * We can move a Return, if it's predecessors are Phi nodes or
166 * comes from another block. In the later case, it is always possible
167 * to move the Return one block up, because the predecessor block must
168 * dominate the Return block (SSA) and then it dominates the predecessor
169 * block of the Return block as well.
171 * All predecessors of the Return block must be Jmp's of course, or we
172 * cannot move it up, so we add blocks if needed.
174 static int can_move_ret(ir_node *ret) {
175 ir_node *retbl = get_nodes_block(ret);
176 int i, n = get_irn_arity(ret);
178 for (i = 0; i < n; ++i) {
179 ir_node *pred = get_irn_n(ret, i);
181 if (! is_Phi(pred) && retbl == get_nodes_block(pred)) {
182 /* first condition failed, found a non-Phi predecessor
183 * then is in the Return block */
188 /* check, that predecessors are Jmps */
189 n = get_Block_n_cfgpreds(retbl);
192 for (i = 0; i < n; ++i) {
193 ir_node *pred = get_Block_cfgpred(retbl, i);
195 pred = skip_Tuple(pred);
196 if (! is_Jmp(pred) && !is_Bad(pred)) {
197 /* simply place a new block here */
198 ir_graph *irg = get_irn_irg(retbl);
199 ir_node *block = new_r_Block(irg, 1, &pred);
200 ir_node *jmp = new_r_Jmp(irg, block);
201 set_Block_cfgpred(retbl, i, jmp);
208 * Normalize the Returns of a graph by moving
209 * the Returns upwards as much as possible.
210 * This might be preferred for code generation.
212 * In pseudocode, it means:
220 * is transformed into
227 void normalize_n_returns(ir_graph *irg) {
228 int i, j, n, n_rets, n_finals, n_ret_vals;
229 ir_node *list = NULL;
230 ir_node *final = NULL;
232 ir_node *endbl = get_irg_end_block(irg);
236 * First, link all returns:
237 * These must be predecessors of the endblock.
238 * Place Returns that can be moved on list, all others
241 n = get_Block_n_cfgpreds(endbl);
242 for (n_finals = n_rets = i = 0; i < n; ++i) {
243 ir_node *ret = get_Block_cfgpred(endbl, i);
245 if (is_Return(ret) && can_move_ret(ret)) {
247 * Ok, all conditions met, we can move this Return, put it
250 set_irn_link(ret, list);
254 /* Put all nodes that are not changed on the final list. */
255 set_irn_link(ret, final);
265 * Now move the Returns upwards. We move always one block up (and create n
266 * new Returns), than we check if a newly created Return can be moved even further.
267 * If yes, we simply add it to our work list, else to the final list.
269 end = get_irg_end(irg);
270 n_ret_vals = get_irn_arity(list);
271 in = alloca(sizeof(*in) * n_ret_vals);
274 ir_node *block = get_nodes_block(ret);
277 list = get_irn_link(ret);
280 n = get_Block_n_cfgpreds(block);
281 for (i = 0; i < n; ++i) {
282 ir_node *jmp = get_Block_cfgpred(block, i);
283 ir_node *new_bl, *new_ret;
289 new_bl = get_nodes_block(jmp);
291 /* create the in-array for the new Return */
292 for (j = 0; j < n_ret_vals; ++j) {
293 ir_node *pred = get_irn_n(ret, j);
295 in[j] = (is_Phi(pred) && get_nodes_block(pred) == block) ? get_Phi_pred(pred, i) : pred;
298 new_ret = new_r_Return(irg, new_bl, in[0], n_ret_vals - 1, &in[1]);
300 if (! is_Bad(new_ret)) {
302 * The newly created node might be bad, if we
303 * create it in a block with only Bad predecessors.
304 * In that case ignore this block.
306 * We could even kill the jmp then ...
308 if (can_move_ret(new_ret)) {
309 set_irn_link(new_ret, list);
313 set_irn_link(new_ret, final);
319 /* remove the Jmp, we have placed a Return here */
320 exchange(jmp, new_r_Bad(irg));
324 * if the memory of the old Return is a PhiM, remove it
325 * from the keep-alives, or it will keep the block which
326 * will crash the dominator algorithm.
328 phiM = get_Return_mem(ret);
330 n = get_End_n_keepalives(end);
331 for (i = 0; i < n; ++i) {
332 if (get_End_keepalive(end, i) == phiM) {
333 set_End_keepalive(end, i, new_r_Bad(irg));
341 * Last step: Create a new endblock, with all nodes on the final
342 * list as predecessors.
344 in = alloca(sizeof(*in) * n_finals);
346 for (i = 0; final; ++i, final = get_irn_link(final))
349 exchange(endbl, new_r_Block(irg, n_finals, in));
351 /* the end block is not automatically skipped, so do it here */
352 set_irg_end_block(irg, skip_Id(get_irg_end_block(irg)));
354 /* Invalidate analysis information:
355 * Blocks become dead and new Returns were deleted, so dominator, outs and loop are inconsistent,
356 * trouts and callee-state should be still valid
358 set_irg_doms_inconsistent(irg);
359 set_irg_extblk_inconsistent(irg); /* may not be needed */
360 set_irg_outs_inconsistent(irg);
361 set_irg_loopinfo_inconsistent(current_ir_graph);