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 prefered 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;
60 ir_node *block, *new_ret;
62 /* look, if we have more than one return */
63 n = get_Block_n_cfgpreds(endbl);
64 returns = alloca((n + 7) >> 3);
65 memset(returns, 0, (n + 7) >> 3);
67 for (n_rets = i = 0; i < n; ++i) {
68 ir_node *node = get_Block_cfgpred(endbl, i);
70 if (get_irn_op(node) == op_Return) {
76 n_ret_vals = get_irn_arity(node);
80 /* there should be at least one Return node in Firm */
84 in = alloca(sizeof(*in) * IMAX(n_rets, n_ret_vals));
85 retvals = alloca(sizeof(*in) * n_rets * n_ret_vals);
87 for (j = i = 0; i < n; ++i) {
89 ir_node *ret = get_Block_cfgpred(endbl, i);
90 ir_node *block = get_nodes_block(ret);
92 /* create a new Jmp for every Ret and place the in in */
93 in[j] = new_r_Jmp(irg, block);
95 /* save the return values and shuffle them */
96 for (k = 0; k < n_ret_vals; ++k)
97 retvals[j + k*n_rets] = get_irn_n(ret, k);
99 set_Block_cfgpred(endbl, i, new_r_Bad(irg));
106 /* ok, create a new block with all created in's */
107 block = new_r_Block(irg, n_rets, in);
109 /* now create the Phi nodes */
110 for (j = i = 0; i < n_ret_vals; ++i, j += n_rets) {
111 /* the return values are already shuffled */
112 in[i] = new_r_Phi(irg, block, n_rets, &retvals[j], get_irn_mode(retvals[j]));
115 new_ret = new_r_Return(irg, block, in[0], n_ret_vals-1, &in[1]);
117 set_Block_cfgpred(endbl, last_idx, new_ret);
119 /* invalidate analysis information:
120 * a new Block was added, so dominator, outs and loop are inconsistent,
121 * trouts and callee-state should be still valid
123 set_irg_dom_inconsistent(irg);
124 set_irg_outs_inconsistent(irg);
125 set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent);
129 * check, whether a Ret can be moved on block upwards.
131 * In a block with a Return, all live nodes must be linked
132 * with the Return, otherwise they are dead (because the Return leaves
133 * the graph, so no more users of the other nodes can exists.
135 * We can move a Return, if it's predecessors are Phi nodes or
136 * comes from another block. In the later case, it is always possible
137 * to move the Return one block up, because the predecessor block must
138 * dominate the Return block (SSA) and then it dominates the predecessor
139 * block of the Return block as well.
141 * All predecessors of the Return block must be Jmp's of course, or we
142 * cannot move it up, so we check this either.
144 static int can_move_ret(ir_node *ret)
146 ir_node *retbl = get_nodes_block(ret);
147 int i, n = get_irn_arity(ret);
149 for (i = 0; i < n; ++i) {
150 ir_node *pred = get_irn_n(ret, i);
152 if (! is_Phi(pred) && retbl == get_nodes_block(pred)) {
153 /* first condition failed, found a non-Phi predecessor
154 * then is in the Return block */
159 /* check, that predecessors are Jmps */
160 n = get_Block_n_cfgpreds(retbl);
161 for (i = 0; i < n; ++i)
162 if (get_irn_op(get_Block_cfgpred(retbl, i)) != op_Jmp)
165 /* if we have 0 control flow predecessors, we cannot move :-) */
170 * Normalize the Returns of a graph by moving
171 * the Returns upwards as much as possible.
172 * This might be prefered for code generation.
174 * In pseudocode, it means:
182 * is transformed into
189 void normalize_n_returns(ir_graph *irg)
191 int i, j, n, n_rets, n_finals, n_ret_vals;
192 ir_node *list = NULL;
193 ir_node *final = NULL;
195 ir_node *endbl = get_irg_end_block(irg);
198 /* first, link all returns */
199 n = get_Block_n_cfgpreds(endbl);
200 for (n_finals = n_rets = i = 0; i < n; ++i) {
201 ir_node *ret = get_Block_cfgpred(endbl, i);
203 if (get_irn_op(ret) == op_Return && can_move_ret(ret)) {
205 * Ok, all conditions met, we can move this Return, put it
208 set_irn_link(ret, list);
213 /* Put all nodes that are not changed on the final list. */
214 set_irn_link(ret, final);
224 * Now move the Returns upwards. We move always one block up (and create n
225 * new Returns), than we check if a newly created Return can be moved even further.
226 * If yes, we simply add it to our work list, else to the final list.
228 end = get_irg_end(irg);
229 n_ret_vals = get_irn_arity(list);
230 in = alloca(sizeof(*in) * n_ret_vals);
233 ir_node *block = get_nodes_block(ret);
236 list = get_irn_link(ret);
239 n = get_Block_n_cfgpreds(block);
240 for (i = 0; i < n; ++i) {
241 ir_node *jmp = get_Block_cfgpred(block, i);
242 ir_node *new_bl = get_nodes_block(jmp);
245 /* create the in-array for the new Ret */
246 for (j = 0; j < n_ret_vals; ++j) {
247 ir_node *pred = get_irn_n(ret, j);
249 in[j] = is_Phi(pred) ? get_Phi_pred(pred, i) : pred;
252 new_ret = new_r_Return(irg, new_bl, in[0], n_ret_vals - 1, &in[1]);
254 if (! is_Bad(new_ret)) {
256 * The newly created node might be bad, if we
257 * create it in a block with only Bad predecessors.
258 * In that case ignore this block.
260 * We could even kill the jmp then ...
262 if (can_move_ret(new_ret)) {
263 set_irn_link(new_ret, list);
268 set_irn_link(new_ret, final);
276 * if the memory of the old Return is a PhiM, remove it
277 * from the keep-alives, or it will keep the block which
278 * will crash the dominator algorithm.
280 phiM = get_Return_mem(ret);
282 n = get_End_n_keepalives(end);
283 for (i = 0; i < n; ++i) {
284 if (get_End_keepalive(end, i) == phiM) {
285 set_End_keepalive(end, i, new_r_Bad(irg));
293 * Last step: Create a new endblock, with all nodes on the final
294 * list as predecessors.
296 in = alloca(sizeof(*in) * n_finals);
298 for (i = 0; final; ++i, final = get_irn_link(final))
301 exchange(endbl, new_r_Block(irg, n_finals, in));
303 /* the end block is not automatically skiped, so do it here */
304 set_irg_end_block(irg, skip_Id(get_irg_end_block(irg)));
306 /* Invalidate analysis information:
307 * Blocks become dead and new Eeturns were deleted, so dominator, outs and loop are inconsistent,
308 * trouts and callee-state should be still valid
310 set_irg_dom_inconsistent(irg);
311 set_irg_outs_inconsistent(irg);
312 set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent);