3 * File name: ir/opt/return.c
4 * Purpose: normalize returns
8 * Copyright: (c) 1998-2005 Universität Karlsruhe
9 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
22 #include "irgraph_t.h"
27 #define set_bit(n) (returns[(n) >> 3] |= 1 << ((n) & 7))
28 #define get_bit(n) (returns[(n) >> 3] & (1 << ((n) & 7)))
31 #define IMAX(a, b) ((a) > (b) ? (a) : (b))
34 * Normalize the Returns of a graph by creating a new End block
35 * with One Return(Phi).
36 * This is the preferred input for the if-conversion.
38 * In pseudocode, it means:
53 void normalize_one_return(ir_graph *irg)
55 ir_node *endbl = get_irg_end_block(irg);
56 int i, j, k, n, last_idx, n_rets, n_ret_vals = -1;
57 unsigned char *returns;
58 ir_node **in, **retvals, **endbl_in;
62 /* look, if we have more than one return */
63 n = get_Block_n_cfgpreds(endbl);
65 /* The end block has no predecessors, we have an endless
66 loop. In that case, no returns exists. */
70 returns = alloca((n + 7) >> 3);
71 memset(returns, 0, (n + 7) >> 3);
73 for (n_rets = i = 0; i < n; ++i) {
74 ir_node *node = get_Block_cfgpred(endbl, i);
76 if (is_Return(node)) {
82 n_ret_vals = get_irn_arity(node);
86 /* there should be at least one Return node in Firm */
90 in = alloca(sizeof(*in) * IMAX(n_rets, n_ret_vals));
91 retvals = alloca(sizeof(*retvals) * n_rets * n_ret_vals);
92 endbl_in = alloca(sizeof(*endbl_in) * n);
95 for (j = i = 0; i < n; ++i) {
96 ir_node *ret = get_Block_cfgpred(endbl, i);
99 ir_node *block = get_nodes_block(ret);
101 /* create a new Jmp for every Ret and place the in in */
102 in[j] = new_r_Jmp(irg, block);
104 /* save the return values and shuffle them */
105 for (k = 0; k < n_ret_vals; ++k)
106 retvals[j + k*n_rets] = get_irn_n(ret, k);
111 endbl_in[last_idx++] = ret;
114 /* ok, create a new block with all created in's */
115 block = new_r_Block(irg, n_rets, in);
117 /* now create the Phi nodes */
118 for (j = i = 0; i < n_ret_vals; ++i, j += n_rets) {
121 /* the return values are already shuffled */
123 /* Beware: normally the Phi constructor automatically replaces a Phi(a,...a) into a
124 but NOT, if a is Unknown. Here, we known that this case can be optimize also,
126 first = retvals[j + 0];
127 for (k = 1; k < n_rets; ++k) {
128 if (retvals[j + k] != first) {
136 in[i] = new_r_Phi(irg, block, n_rets, &retvals[j], get_irn_mode(retvals[j]));
139 endbl_in[last_idx++] = new_r_Return(irg, block, in[0], n_ret_vals-1, &in[1]);
141 set_irn_in(endbl, last_idx, endbl_in);
143 /* invalidate analysis information:
144 * a new Block was added, so dominator, outs and loop are inconsistent,
145 * trouts and callee-state should be still valid
147 set_irg_doms_inconsistent(irg);
148 set_irg_outs_inconsistent(irg);
149 set_irg_extblk_inconsistent(irg);
150 set_irg_loopinfo_state(irg, loopinfo_cf_inconsistent);
154 * check, whether a Ret can be moved on block upwards.
156 * In a block with a Return, all live nodes must be linked
157 * with the Return, otherwise they are dead (because the Return leaves
158 * the graph, so no more users of the other nodes can exists.
160 * We can move a Return, if it's predecessors are Phi nodes or
161 * comes from another block. In the later case, it is always possible
162 * to move the Return one block up, because the predecessor block must
163 * dominate the Return block (SSA) and then it dominates the predecessor
164 * block of the Return block as well.
166 * All predecessors of the Return block must be Jmp's of course, or we
167 * cannot move it up, so we check this either.
169 static int can_move_ret(ir_node *ret)
171 ir_node *retbl = get_nodes_block(ret);
172 int i, n = get_irn_arity(ret);
174 for (i = 0; i < n; ++i) {
175 ir_node *pred = get_irn_n(ret, i);
177 if (! is_Phi(pred) && retbl == get_nodes_block(pred)) {
178 /* first condition failed, found a non-Phi predecessor
179 * then is in the Return block */
184 /* check, that predecessors are Jmps */
185 n = get_Block_n_cfgpreds(retbl);
186 for (i = 0; i < n; ++i)
187 if (get_irn_op(get_Block_cfgpred(retbl, i)) != op_Jmp)
190 /* if we have 0 control flow predecessors, we cannot move :-) */
195 * Normalize the Returns of a graph by moving
196 * the Returns upwards as much as possible.
197 * This might be preferred for code generation.
199 * In pseudocode, it means:
207 * is transformed into
214 void normalize_n_returns(ir_graph *irg)
216 int i, j, n, n_rets, n_finals, n_ret_vals;
217 ir_node *list = NULL;
218 ir_node *final = NULL;
220 ir_node *endbl = get_irg_end_block(irg);
224 * First, link all returns:
225 * These must be predecessors of the endblock.
226 * Place Returns that can be moved on list, all others
229 n = get_Block_n_cfgpreds(endbl);
230 for (n_finals = n_rets = i = 0; i < n; ++i) {
231 ir_node *ret = get_Block_cfgpred(endbl, i);
233 if (is_Return(ret) && can_move_ret(ret)) {
235 * Ok, all conditions met, we can move this Return, put it
238 set_irn_link(ret, list);
243 /* Put all nodes that are not changed on the final list. */
244 set_irn_link(ret, final);
254 * Now move the Returns upwards. We move always one block up (and create n
255 * new Returns), than we check if a newly created Return can be moved even further.
256 * If yes, we simply add it to our work list, else to the final list.
258 end = get_irg_end(irg);
259 n_ret_vals = get_irn_arity(list);
260 in = alloca(sizeof(*in) * n_ret_vals);
263 ir_node *block = get_nodes_block(ret);
266 list = get_irn_link(ret);
269 n = get_Block_n_cfgpreds(block);
270 for (i = 0; i < n; ++i) {
271 ir_node *jmp = get_Block_cfgpred(block, i);
272 ir_node *new_bl, *new_ret;
274 if (get_irn_op(jmp) != op_Jmp)
277 new_bl = get_nodes_block(jmp);
279 /* create the in-array for the new Ret */
280 for (j = 0; j < n_ret_vals; ++j) {
281 ir_node *pred = get_irn_n(ret, j);
283 in[j] = (is_Phi(pred) && get_nodes_block(pred) == block) ? get_Phi_pred(pred, i) : pred;
286 new_ret = new_r_Return(irg, new_bl, in[0], n_ret_vals - 1, &in[1]);
288 if (! is_Bad(new_ret)) {
290 * The newly created node might be bad, if we
291 * create it in a block with only Bad predecessors.
292 * In that case ignore this block.
294 * We could even kill the jmp then ...
296 if (can_move_ret(new_ret)) {
297 set_irn_link(new_ret, list);
302 set_irn_link(new_ret, final);
308 /* remove the Jmp, we have placed a Return here */
309 exchange(jmp, new_r_Bad(irg));
313 * if the memory of the old Return is a PhiM, remove it
314 * from the keep-alives, or it will keep the block which
315 * will crash the dominator algorithm.
317 phiM = get_Return_mem(ret);
319 n = get_End_n_keepalives(end);
320 for (i = 0; i < n; ++i) {
321 if (get_End_keepalive(end, i) == phiM) {
322 set_End_keepalive(end, i, new_r_Bad(irg));
330 * Last step: Create a new endblock, with all nodes on the final
331 * list as predecessors.
333 in = alloca(sizeof(*in) * n_finals);
335 for (i = 0; final; ++i, final = get_irn_link(final))
338 exchange(endbl, new_r_Block(irg, n_finals, in));
340 /* the end block is not automatically skipped, so do it here */
341 set_irg_end_block(irg, skip_Id(get_irg_end_block(irg)));
343 /* Invalidate analysis information:
344 * Blocks become dead and new Returns were deleted, so dominator, outs and loop are inconsistent,
345 * trouts and callee-state should be still valid
347 set_irg_doms_inconsistent(irg);
348 set_irg_outs_inconsistent(irg);
349 set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent);