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 Helper functions for handling ABI constraints in the code
24 * @author Matthias Braun
29 #include "beabihelper.h"
36 #include "irphase_t.h"
39 typedef struct reg_flag_t {
40 const arch_register_t *reg; /**< register at an input position.
41 may be NULL in case of memory input */
42 arch_irn_flags_t flags;
46 * A register state mapping keeps track of the symbol values (=firm nodes)
47 * to registers. This is usefull when constructing straight line code
48 * which like the function prolog or epilog in some architectures.
50 typedef struct register_state_mapping_t {
51 ir_node **value_map; /**< mapping of state indices to values */
52 int **reg_index_map; /**< mapping of regclass,regnum to an index
54 reg_flag_t *regs; /**< registers (and memory values) that form a
56 ir_node *last_barrier;
57 } register_state_mapping_t;
59 struct beabi_helper_env_t {
61 register_state_mapping_t prolog;
62 register_state_mapping_t epilog;
63 ir_phase *stack_order;
66 static void prepare_rsm(register_state_mapping_t *rsm,
67 const arch_env_t *arch_env)
69 unsigned n_reg_classes = arch_env_get_n_reg_class(arch_env);
71 reg_flag_t memory = { NULL, 0 };
73 rsm->regs = NEW_ARR_F(reg_flag_t, 0);
74 /* memory input at 0 */
75 ARR_APP1(reg_flag_t, rsm->regs, memory);
77 rsm->value_map = NULL;
78 rsm->reg_index_map = XMALLOCN(int*, n_reg_classes);
79 for (c = 0; c < n_reg_classes; ++c) {
80 const arch_register_class_t *cls = arch_env_get_reg_class(arch_env, c);
81 unsigned n_regs = arch_register_class_n_regs(cls);
84 rsm->reg_index_map[c] = XMALLOCN(int, n_regs);
85 for (r = 0; r < n_regs; ++r) {
86 rsm->reg_index_map[c][r] = -1;
91 static void free_rsm(register_state_mapping_t *rsm, const arch_env_t *arch_env)
93 unsigned n_reg_classes = arch_env_get_n_reg_class(arch_env);
96 for (c = 0; c < n_reg_classes; ++c) {
97 free(rsm->reg_index_map[c]);
100 free(rsm->reg_index_map);
101 if (rsm->value_map != NULL)
102 DEL_ARR_F(rsm->value_map);
103 DEL_ARR_F(rsm->regs);
106 rsm->reg_index_map = NULL;
107 rsm->value_map = NULL;
110 static void rsm_clear_regs(register_state_mapping_t *rsm,
111 const arch_env_t *arch_env)
113 unsigned n_reg_classes = arch_env_get_n_reg_class(arch_env);
115 reg_flag_t memory = { NULL, 0 };
117 for (c = 0; c < n_reg_classes; ++c) {
118 const arch_register_class_t *cls = arch_env_get_reg_class(arch_env, c);
119 unsigned n_regs = arch_register_class_n_regs(cls);
122 for (r = 0; r < n_regs; ++r) {
123 rsm->reg_index_map[c][r] = -1;
126 ARR_RESIZE(reg_flag_t, rsm->regs, 0);
127 ARR_APP1(reg_flag_t, rsm->regs, memory);
129 if (rsm->value_map != NULL) {
130 DEL_ARR_F(rsm->value_map);
131 rsm->value_map = NULL;
135 static int rsm_add_reg(register_state_mapping_t *rsm,
136 const arch_register_t *reg, arch_irn_flags_t flags)
138 int input_idx = ARR_LEN(rsm->regs);
139 int cls_idx = reg->reg_class->index;
140 int reg_idx = reg->index;
141 reg_flag_t regflag = { reg, flags };
143 /* we must not have used get_value yet */
144 assert(rsm->reg_index_map[cls_idx][reg_idx] == -1);
145 rsm->reg_index_map[cls_idx][reg_idx] = input_idx;
146 ARR_APP1(reg_flag_t, rsm->regs, regflag);
148 if (rsm->value_map != NULL) {
149 ARR_APP1(ir_node*, rsm->value_map, NULL);
150 assert(ARR_LEN(rsm->value_map) == ARR_LEN(rsm->regs));
156 static ir_node *rsm_get_value(register_state_mapping_t *rsm, int index)
158 assert(0 <= index && index < ARR_LEN(rsm->value_map));
159 return rsm->value_map[index];
162 static ir_node *rsm_get_reg_value(register_state_mapping_t *rsm,
163 const arch_register_t *reg)
165 int cls_idx = reg->reg_class->index;
166 int reg_idx = reg->index;
167 int input_idx = rsm->reg_index_map[cls_idx][reg_idx];
169 return rsm_get_value(rsm, input_idx);
172 static void rsm_set_value(register_state_mapping_t *rsm, int index,
175 assert(0 <= index && index < ARR_LEN(rsm->value_map));
176 rsm->value_map[index] = value;
179 static void rsm_set_reg_value(register_state_mapping_t *rsm,
180 const arch_register_t *reg, ir_node *value)
182 int cls_idx = reg->reg_class->index;
183 int reg_idx = reg->index;
184 int input_idx = rsm->reg_index_map[cls_idx][reg_idx];
185 rsm_set_value(rsm, input_idx, value);
188 static ir_node *rsm_create_barrier(register_state_mapping_t *rsm,
191 int n_barrier_outs = ARR_LEN(rsm->regs);
192 ir_node **in = rsm->value_map;
196 assert(ARR_LEN(rsm->value_map) == n_barrier_outs);
198 barrier = be_new_Barrier(block, n_barrier_outs, in);
200 for (o = 0; o < n_barrier_outs; ++o) {
201 const reg_flag_t *regflag = &rsm->regs[o];
202 const arch_register_t *reg = regflag->reg;
205 arch_set_out_register_req(barrier, o, arch_no_register_req);
206 proj = new_r_Proj(barrier, mode_M, o);
208 be_set_constr_single_reg_in(barrier, o, reg, 0);
209 be_set_constr_single_reg_out(barrier, o, reg, regflag->flags);
210 proj = new_r_Proj(barrier, reg->reg_class->mode, o);
212 rsm->value_map[o] = proj;
215 rsm->last_barrier = barrier;
224 beabi_helper_env_t *be_abihelper_prepare(ir_graph *irg)
226 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
227 beabi_helper_env_t *env = XMALLOCZ(beabi_helper_env_t);
230 prepare_rsm(&env->prolog, arch_env);
231 prepare_rsm(&env->epilog, arch_env);
236 void be_abihelper_finish(beabi_helper_env_t *env)
238 const arch_env_t *arch_env = be_get_irg_arch_env(env->irg);
240 free_rsm(&env->prolog, arch_env);
241 if (env->epilog.reg_index_map != NULL) {
242 free_rsm(&env->epilog, arch_env);
247 void be_prolog_add_reg(beabi_helper_env_t *env, const arch_register_t *reg,
248 arch_irn_flags_t flags)
250 rsm_add_reg(&env->prolog, reg, flags);
253 ir_node *be_prolog_create_start(beabi_helper_env_t *env, dbg_info *dbgi,
256 int n_start_outs = ARR_LEN(env->prolog.regs);
257 ir_node *start = be_new_Start(dbgi, block, n_start_outs);
260 assert(env->prolog.value_map == NULL);
261 env->prolog.value_map = NEW_ARR_F(ir_node*, n_start_outs);
263 for (o = 0; o < n_start_outs; ++o) {
264 const reg_flag_t *regflag = &env->prolog.regs[o];
265 const arch_register_t *reg = regflag->reg;
268 arch_set_out_register_req(start, o, arch_no_register_req);
269 proj = new_r_Proj(start, mode_M, o);
271 be_set_constr_single_reg_out(start, o, regflag->reg,
273 arch_irn_set_register(start, o, regflag->reg);
274 proj = new_r_Proj(start, reg->reg_class->mode, o);
276 env->prolog.value_map[o] = proj;
279 /* start node should really be the first thing constructed */
280 assert(env->prolog.last_barrier == NULL);
281 env->prolog.last_barrier = start;
286 ir_node *be_prolog_create_barrier(beabi_helper_env_t *env, ir_node *block)
288 return rsm_create_barrier(&env->prolog, block);
291 ir_node *be_prolog_get_reg_value(beabi_helper_env_t *env,
292 const arch_register_t *reg)
294 return rsm_get_reg_value(&env->prolog, reg);
297 ir_node *be_prolog_get_memory(beabi_helper_env_t *env)
299 return rsm_get_value(&env->prolog, 0);
302 void be_prolog_set_reg_value(beabi_helper_env_t *env,
303 const arch_register_t *reg, ir_node *value)
305 rsm_set_reg_value(&env->prolog, reg, value);
308 void be_prolog_set_memory(beabi_helper_env_t *env, ir_node *value)
310 rsm_set_value(&env->prolog, 0, value);
315 void be_epilog_begin(beabi_helper_env_t *env)
317 const arch_env_t *arch_env = be_get_irg_arch_env(env->irg);
318 rsm_clear_regs(&env->epilog, arch_env);
319 env->epilog.value_map = NEW_ARR_F(ir_node*, 1);
320 env->epilog.value_map[0] = NULL;
323 void be_epilog_add_reg(beabi_helper_env_t *env, const arch_register_t *reg,
324 arch_irn_flags_t flags, ir_node *value)
326 int index = rsm_add_reg(&env->epilog, reg, flags);
327 rsm_set_value(&env->epilog, index, value);
330 void be_epilog_set_reg_value(beabi_helper_env_t *env,
331 const arch_register_t *reg, ir_node *value)
333 rsm_set_reg_value(&env->epilog, reg, value);
336 void be_epilog_set_memory(beabi_helper_env_t *env, ir_node *value)
338 rsm_set_value(&env->epilog, 0, value);
341 ir_node *be_epilog_get_reg_value(beabi_helper_env_t *env,
342 const arch_register_t *reg)
344 return rsm_get_reg_value(&env->epilog, reg);
347 ir_node *be_epilog_get_memory(beabi_helper_env_t *env)
349 return rsm_get_value(&env->epilog, 0);
352 ir_node *be_epilog_create_barrier(beabi_helper_env_t *env, ir_node *block)
354 return rsm_create_barrier(&env->epilog, block);
357 ir_node *be_epilog_create_return(beabi_helper_env_t *env, dbg_info *dbgi,
360 int n_return_in = ARR_LEN(env->epilog.regs);
361 ir_node **in = env->epilog.value_map;
362 int n_res = 1; /* TODO */
363 unsigned pop = 0; /* TODO */
367 assert(ARR_LEN(env->epilog.value_map) == n_return_in);
369 ret = be_new_Return(dbgi, get_irn_irg(block), block, n_res, pop,
371 for (i = 0; i < n_return_in; ++i) {
372 const reg_flag_t *regflag = &env->epilog.regs[i];
373 const arch_register_t *reg = regflag->reg;
375 be_set_constr_single_reg_in(ret, i, reg, 0);
379 rsm_clear_regs(&env->epilog, be_get_irg_arch_env(env->irg));
380 env->epilog.last_barrier = NULL;
385 static void add_missing_keep_walker(ir_node *node, void *data)
388 unsigned *found_projs;
389 const ir_edge_t *edge;
390 ir_mode *mode = get_irn_mode(node);
396 n_outs = arch_irn_get_n_outs(node);
400 rbitset_alloca(found_projs, n_outs);
401 foreach_out_edge(node, edge) {
402 ir_node *succ = get_edge_src_irn(edge);
405 /* The node could be kept */
406 if (is_End(succ) || is_Anchor(succ))
409 if (get_irn_mode(succ) == mode_M)
412 pn = get_Proj_proj(succ);
414 rbitset_set(found_projs, pn);
418 /* are keeps missing? */
420 for (i = 0; i < n_outs; ++i) {
423 const arch_register_req_t *req;
424 const arch_register_class_t *cls;
426 if (rbitset_is_set(found_projs, i)) {
430 req = arch_get_out_register_req(node, i);
432 if (cls == NULL || (cls->flags & arch_register_class_flag_manual_ra)) {
436 block = get_nodes_block(node);
437 in[0] = new_r_Proj(node, arch_register_class_mode(cls), i);
438 if (last_keep != NULL) {
439 be_Keep_add_node(last_keep, cls, in[0]);
441 last_keep = be_new_Keep(block, 1, in);
442 if (sched_is_scheduled(node)) {
443 sched_add_after(node, last_keep);
449 void be_add_missing_keeps(ir_graph *irg)
451 irg_walk_graph(irg, add_missing_keep_walker, NULL, NULL);
456 static void collect_node(ir_node *node)
458 ir_node *block = get_nodes_block(node);
459 ir_node *old = get_irn_link(block);
461 set_irn_link(node, old);
462 set_irn_link(block, node);
465 static void link_ops_in_block_walker(ir_node *node, void *data)
469 switch (get_irn_opcode(node)) {
475 /** all non-stack alloc nodes should be lowered before the backend */
476 assert(get_Alloc_where(node) == stack_alloc);
480 assert(get_Free_where(node) == stack_alloc);
484 if (get_Builtin_kind(node) == ir_bk_return_address) {
485 ir_node *param = get_Builtin_param(node, 0);
486 tarval *tv = get_Const_tarval(param); /* must be Const */
487 long value = get_tarval_long(tv);
489 /* we need esp for the climbframe algo */
499 static heights_t *heights;
502 * Check if a node is somehow data dependent on another one.
503 * both nodes must be in the same basic block.
504 * @param n1 The first node.
505 * @param n2 The second node.
506 * @return 1, if n1 is data dependent (transitively) on n2, 0 if not.
508 static int dependent_on(const ir_node *n1, const ir_node *n2)
510 assert(get_nodes_block(n1) == get_nodes_block(n2));
512 return heights_reachable_in_block(heights, n1, n2);
515 static int cmp_call_dependency(const void *c1, const void *c2)
517 const ir_node *n1 = *(const ir_node **) c1;
518 const ir_node *n2 = *(const ir_node **) c2;
521 Classical qsort() comparison function behavior:
522 0 if both elements are equal
523 1 if second is "smaller" that first
524 -1 if first is "smaller" that second
526 if (dependent_on(n1, n2))
529 if (dependent_on(n2, n1))
532 /* The nodes have no depth order, but we need a total order because qsort()
534 return get_irn_idx(n2) - get_irn_idx(n1);
537 static void process_ops_in_block(ir_node *block, void *data)
539 ir_phase *phase = data;
546 for (node = get_irn_link(block); node != NULL; node = get_irn_link(node)) {
553 nodes = XMALLOCN(ir_node*, n_nodes);
555 for (node = get_irn_link(block); node != NULL; node = get_irn_link(node)) {
558 assert(n == n_nodes);
560 /* order nodes according to their data dependencies */
561 qsort(nodes, n_nodes, sizeof(nodes[0]), cmp_call_dependency);
563 for (n = n_nodes-1; n > 0; --n) {
564 ir_node *node = nodes[n];
565 ir_node *pred = nodes[n-1];
567 phase_set_irn_data(phase, node, pred);
571 void be_collect_stacknodes(beabi_helper_env_t *env)
573 ir_graph *irg = env->irg;
574 irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, NULL);
576 assert(env->stack_order == NULL);
577 env->stack_order = new_phase(irg, phase_irn_init_default);
579 heights = heights_new(irg);
580 irg_block_walk_graph(irg, NULL, process_ops_in_block, env->stack_order);
581 heights_free(heights);
584 ir_node *be_get_stack_pred(const beabi_helper_env_t *env, const ir_node *node)
586 return phase_get_irn_data(env->stack_order, node);