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
28 #include "beabihelper.h"
35 #include "irnodemap.h"
40 * An entry in the register state map.
42 typedef struct reg_flag_t {
43 const arch_register_t *reg; /**< register at an input position.
44 may be NULL in case of memory input */
45 arch_register_req_type_t flags; /**< requirement flags for this register. */
49 * A register state mapping keeps track of the symbol values (=firm nodes)
50 * to registers. This is useful when constructing straight line code
51 * like the function prolog or epilog in some architectures.
53 typedef struct register_state_mapping_t {
54 ir_node **value_map; /**< mapping of state indices to values */
55 size_t **reg_index_map; /**< mapping of regclass,regnum to an index
57 reg_flag_t *regs; /**< registers (and memory values) that form a
59 } register_state_mapping_t;
62 * The environment for all helper functions.
64 struct beabi_helper_env_t {
65 ir_graph *irg; /**< the graph we operate on */
66 register_state_mapping_t prolog; /**< the register state map for the prolog */
67 register_state_mapping_t epilog; /**< the register state map for the epilog */
71 * Create a new empty register state map for the given
74 * @param rsm the register state map to be initialized
75 * @param arch_env the architecture environment
77 * After this call, the register map is initialized to empty.
79 static void prepare_rsm(register_state_mapping_t *rsm,
80 const arch_env_t *arch_env)
82 unsigned n_reg_classes = arch_env->n_register_classes;
84 reg_flag_t memory = { NULL, arch_register_req_type_none };
86 rsm->regs = NEW_ARR_F(reg_flag_t, 0);
87 /* memory input at 0 */
88 ARR_APP1(reg_flag_t, rsm->regs, memory);
90 rsm->value_map = NULL;
91 rsm->reg_index_map = XMALLOCN(size_t*, n_reg_classes);
92 for (c = 0; c < n_reg_classes; ++c) {
93 const arch_register_class_t *cls = &arch_env->register_classes[c];
94 unsigned n_regs = arch_register_class_n_regs(cls);
97 rsm->reg_index_map[c] = XMALLOCN(size_t, n_regs);
98 for (r = 0; r < n_regs; ++r) {
99 rsm->reg_index_map[c][r] = (size_t)-1;
105 * Destroy a register state map for the given
108 * @param rsm the register state map to be destroyed
109 * @param arch_env the architecture environment
111 * After this call, the register map is initialized to empty.
113 static void free_rsm(register_state_mapping_t *rsm, const arch_env_t *arch_env)
115 unsigned n_reg_classes = arch_env->n_register_classes;
118 for (c = 0; c < n_reg_classes; ++c) {
119 free(rsm->reg_index_map[c]);
122 free(rsm->reg_index_map);
123 if (rsm->value_map != NULL)
124 DEL_ARR_F(rsm->value_map);
125 DEL_ARR_F(rsm->regs);
128 rsm->reg_index_map = NULL;
129 rsm->value_map = NULL;
133 * Remove all registers from a register state map.
135 * @param rsm the register state map to be destroyed
136 * @param arch_env the architecture environment
138 static void rsm_clear_regs(register_state_mapping_t *rsm,
139 const arch_env_t *arch_env)
141 unsigned n_reg_classes = arch_env->n_register_classes;
143 reg_flag_t memory = { NULL, arch_register_req_type_none };
145 for (c = 0; c < n_reg_classes; ++c) {
146 const arch_register_class_t *cls = &arch_env->register_classes[c];
147 unsigned n_regs = arch_register_class_n_regs(cls);
150 for (r = 0; r < n_regs; ++r) {
151 rsm->reg_index_map[c][r] = (size_t)-1;
154 ARR_RESIZE(reg_flag_t, rsm->regs, 0);
155 ARR_APP1(reg_flag_t, rsm->regs, memory);
157 if (rsm->value_map != NULL) {
158 DEL_ARR_F(rsm->value_map);
159 rsm->value_map = NULL;
164 * Add a register and its constraint flags to a register state map
165 * and return its index inside the map.
167 static size_t rsm_add_reg(register_state_mapping_t *rsm,
168 const arch_register_t *reg,
169 arch_register_req_type_t flags)
171 size_t input_idx = ARR_LEN(rsm->regs);
172 int cls_idx = reg->reg_class->index;
173 int reg_idx = reg->index;
174 reg_flag_t regflag = { reg, flags };
176 /* we must not have used get_value yet */
177 assert(rsm->reg_index_map[cls_idx][reg_idx] == (size_t)-1);
178 rsm->reg_index_map[cls_idx][reg_idx] = input_idx;
179 ARR_APP1(reg_flag_t, rsm->regs, regflag);
181 if (rsm->value_map != NULL) {
182 ARR_APP1(ir_node*, rsm->value_map, NULL);
183 assert(ARR_LEN(rsm->value_map) == ARR_LEN(rsm->regs));
189 * Retrieve the ir_node stored at the given index in the register state map.
191 static ir_node *rsm_get_value(register_state_mapping_t *rsm, size_t index)
193 assert(index < ARR_LEN(rsm->value_map));
194 return rsm->value_map[index];
198 * Retrieve the ir_node occupying the given register in the register state map.
200 static ir_node *rsm_get_reg_value(register_state_mapping_t *rsm,
201 const arch_register_t *reg)
203 int cls_idx = reg->reg_class->index;
204 int reg_idx = reg->index;
205 size_t input_idx = rsm->reg_index_map[cls_idx][reg_idx];
207 return rsm_get_value(rsm, input_idx);
211 * Enter a ir_node at the given index in the register state map.
213 static void rsm_set_value(register_state_mapping_t *rsm, size_t index,
216 assert(index < ARR_LEN(rsm->value_map));
217 rsm->value_map[index] = value;
221 * Enter a ir_node at the given register in the register state map.
223 static void rsm_set_reg_value(register_state_mapping_t *rsm,
224 const arch_register_t *reg, ir_node *value)
226 int cls_idx = reg->reg_class->index;
227 int reg_idx = reg->index;
228 size_t input_idx = rsm->reg_index_map[cls_idx][reg_idx];
229 rsm_set_value(rsm, input_idx, value);
233 beabi_helper_env_t *be_abihelper_prepare(ir_graph *irg)
235 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
236 beabi_helper_env_t *env = XMALLOCZ(beabi_helper_env_t);
239 prepare_rsm(&env->prolog, arch_env);
240 prepare_rsm(&env->epilog, arch_env);
245 void be_abihelper_finish(beabi_helper_env_t *env)
247 const arch_env_t *arch_env = be_get_irg_arch_env(env->irg);
249 free_rsm(&env->prolog, arch_env);
250 if (env->epilog.reg_index_map != NULL) {
251 free_rsm(&env->epilog, arch_env);
256 void be_prolog_add_reg(beabi_helper_env_t *env, const arch_register_t *reg,
257 arch_register_req_type_t flags)
259 rsm_add_reg(&env->prolog, reg, flags);
262 ir_node *be_prolog_create_start(beabi_helper_env_t *env, dbg_info *dbgi,
265 int n_start_outs = ARR_LEN(env->prolog.regs);
266 ir_node *start = be_new_Start(dbgi, block, n_start_outs);
269 assert(env->prolog.value_map == NULL);
270 env->prolog.value_map = NEW_ARR_F(ir_node*, n_start_outs);
272 for (o = 0; o < n_start_outs; ++o) {
273 const reg_flag_t *regflag = &env->prolog.regs[o];
274 const arch_register_t *reg = regflag->reg;
277 arch_set_irn_register_req_out(start, o, arch_no_register_req);
278 proj = new_r_Proj(start, mode_M, o);
280 be_set_constr_single_reg_out(start, o, regflag->reg,
282 arch_set_irn_register_out(start, o, regflag->reg);
283 proj = new_r_Proj(start, reg->reg_class->mode, o);
285 env->prolog.value_map[o] = proj;
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_register_req_type_t flags, ir_node *value)
326 size_t 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_return(beabi_helper_env_t *env, dbg_info *dbgi,
355 size_t n_return_in = ARR_LEN(env->epilog.regs);
356 ir_node **in = env->epilog.value_map;
357 int n_res = 1; /* TODO */
358 unsigned pop = 0; /* TODO */
362 assert(ARR_LEN(env->epilog.value_map) == n_return_in);
364 ret = be_new_Return(dbgi, get_irn_irg(block), block, n_res, pop,
366 for (i = 0; i < n_return_in; ++i) {
367 const reg_flag_t *regflag = &env->epilog.regs[i];
368 const arch_register_t *reg = regflag->reg;
370 be_set_constr_single_reg_in(ret, i, reg,
371 arch_register_req_type_none);
375 rsm_clear_regs(&env->epilog, be_get_irg_arch_env(env->irg));
381 * Tests whether a node has a real user and is not just kept by the End or
384 static bool has_real_user(const ir_node *node)
386 foreach_out_edge(node, edge) {
387 ir_node *user = get_edge_src_irn(edge);
388 if (!is_End(user) && !is_Anchor(user))
394 static ir_node *add_to_keep(ir_node *last_keep,
395 const arch_register_class_t *cls, ir_node *node)
397 if (last_keep != NULL) {
398 be_Keep_add_node(last_keep, cls, node);
400 ir_node *in[1] = { node };
401 ir_node *block = get_nodes_block(node);
403 last_keep = be_new_Keep(block, 1, in);
405 schedpoint = skip_Proj(node);
406 if (sched_is_scheduled(schedpoint)) {
407 sched_add_after(schedpoint, last_keep);
413 void be_add_missing_keeps_node(ir_node *node)
416 unsigned *found_projs;
417 ir_mode *mode = get_irn_mode(node);
419 ir_node **existing_projs;
421 if (mode != mode_T) {
422 if (!has_real_user(node)) {
423 const arch_register_req_t *req = arch_get_irn_register_req(node);
424 const arch_register_class_t *cls = req->cls;
426 || (cls->flags & arch_register_class_flag_manual_ra)) {
430 add_to_keep(NULL, cls, node);
435 n_outs = arch_get_irn_n_outs(node);
439 rbitset_alloca(found_projs, n_outs);
440 existing_projs = ALLOCANZ(ir_node*, n_outs);
441 foreach_out_edge(node, edge) {
442 ir_node *succ = get_edge_src_irn(edge);
443 ir_mode *mode = get_irn_mode(succ);
446 /* The node could be kept */
447 if (is_End(succ) || is_Anchor(succ))
449 if (mode == mode_M || mode == mode_X)
451 pn = get_Proj_proj(succ);
452 existing_projs[pn] = succ;
453 if (!has_real_user(succ))
457 rbitset_set(found_projs, pn);
460 /* are keeps missing? */
462 for (i = 0; i < n_outs; ++i) {
464 const arch_register_req_t *req;
465 const arch_register_class_t *cls;
467 if (rbitset_is_set(found_projs, i)) {
471 req = arch_get_irn_register_req_out(node, i);
473 if (cls == NULL || (cls->flags & arch_register_class_flag_manual_ra)) {
477 value = existing_projs[i];
479 value = new_r_Proj(node, arch_register_class_mode(cls), i);
480 last_keep = add_to_keep(last_keep, cls, value);
484 static void add_missing_keep_walker(ir_node *node, void *data)
487 be_add_missing_keeps_node(node);
490 void be_add_missing_keeps(ir_graph *irg)
492 irg_walk_graph(irg, add_missing_keep_walker, NULL, NULL);
497 * Link the node into its block list as a new head.
499 static void collect_node(ir_node *node)
501 ir_node *block = get_nodes_block(node);
502 ir_node *old = (ir_node*)get_irn_link(block);
504 set_irn_link(node, old);
505 set_irn_link(block, node);
509 * Post-walker: link all nodes that probably access the stack into lists of their block.
511 static void link_ops_in_block_walker(ir_node *node, void *data)
515 switch (get_irn_opcode(node)) {
521 /** all non-stack alloc nodes should be lowered before the backend */
522 assert(get_Alloc_where(node) == stack_alloc);
526 assert(get_Free_where(node) == stack_alloc);
530 if (get_Builtin_kind(node) == ir_bk_return_address) {
531 ir_node *param = get_Builtin_param(node, 0);
532 ir_tarval *tv = get_Const_tarval(param); /* must be Const */
533 long value = get_tarval_long(tv);
535 /* not the return address of the current function:
536 * we need the stack pointer for the frame climbing */
546 static ir_heights_t *heights;
549 * Check if a node is somehow data dependent on another one.
550 * both nodes must be in the same basic block.
551 * @param n1 The first node.
552 * @param n2 The second node.
553 * @return 1, if n1 is data dependent (transitively) on n2, 0 if not.
555 static int dependent_on(const ir_node *n1, const ir_node *n2)
557 assert(get_nodes_block(n1) == get_nodes_block(n2));
559 return heights_reachable_in_block(heights, n1, n2);
563 * Classical qsort() comparison function behavior:
565 * 0 if both elements are equal, no node depend on the other
566 * +1 if first depends on second (first is greater)
567 * -1 if second depends on first (second is greater)
569 static int cmp_call_dependency(const void *c1, const void *c2)
571 const ir_node *n1 = *(const ir_node **) c1;
572 const ir_node *n2 = *(const ir_node **) c2;
575 if (dependent_on(n1, n2))
578 if (dependent_on(n2, n1))
581 /* The nodes have no depth order, but we need a total order because qsort()
584 * Additionally, we need to respect transitive dependencies. Consider a
585 * Call a depending on Call b and an independent Call c.
586 * We MUST NOT order c > a and b > c. */
587 h1 = get_irn_height(heights, n1);
588 h2 = get_irn_height(heights, n2);
589 if (h1 < h2) return 1;
590 if (h1 > h2) return -1;
591 /* Same height, so use a random (but stable) order */
592 return get_irn_idx(n2) - get_irn_idx(n1);
596 * Block-walker: sorts dependencies and remember them into a phase
598 static void process_ops_in_block(ir_node *block, void *data)
600 ir_nodemap *map = (ir_nodemap*)data;
607 for (node = (ir_node*)get_irn_link(block); node != NULL;
608 node = (ir_node*)get_irn_link(node)) {
615 nodes = XMALLOCN(ir_node*, n_nodes);
617 for (node = (ir_node*)get_irn_link(block); node != NULL;
618 node = (ir_node*)get_irn_link(node)) {
621 assert(n == n_nodes);
623 /* order nodes according to their data dependencies */
624 qsort(nodes, n_nodes, sizeof(nodes[0]), cmp_call_dependency);
626 /* remember the calculated dependency into a phase */
627 for (n = n_nodes-1; n > 0; --n) {
628 ir_node *node = nodes[n];
629 ir_node *pred = nodes[n-1];
631 ir_nodemap_insert(map, node, pred);
638 struct be_stackorder_t {
639 ir_nodemap stack_order; /**< a phase to handle stack dependencies. */
642 be_stackorder_t *be_collect_stacknodes(ir_graph *irg)
644 be_stackorder_t *env = XMALLOCZ(be_stackorder_t);
646 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
648 /* collect all potential^stack accessing nodes */
649 irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, NULL);
651 ir_nodemap_init(&env->stack_order, irg);
653 /* use heights to create a total order for those nodes: this order is stored
654 * in the created phase */
655 heights = heights_new(irg);
656 irg_block_walk_graph(irg, NULL, process_ops_in_block, &env->stack_order);
657 heights_free(heights);
659 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
664 ir_node *be_get_stack_pred(const be_stackorder_t *env, const ir_node *node)
666 return ir_nodemap_get(ir_node, &env->stack_order, node);
669 void be_free_stackorder(be_stackorder_t *env)
671 ir_nodemap_destroy(&env->stack_order);
675 static void create_stores_for_type(ir_graph *irg, ir_type *type)
677 size_t n = get_compound_n_members(type);
678 ir_node *frame = get_irg_frame(irg);
679 ir_node *initial_mem = get_irg_initial_mem(irg);
680 ir_node *mem = initial_mem;
681 ir_node *first_store = NULL;
682 ir_node *start_block = get_irg_start_block(irg);
683 ir_node *args = get_irg_args(irg);
686 /* all parameter entities left in the frame type require stores.
687 * (The ones passed on the stack have been moved to the arg type) */
688 for (i = 0; i < n; ++i) {
689 ir_entity *entity = get_compound_member(type, i);
692 if (!is_parameter_entity(entity))
695 arg = get_entity_parameter_number(entity);
696 if (arg == IR_VA_START_PARAMETER_NUMBER)
699 addr = new_r_Sel(start_block, mem, frame, 0, NULL, entity);
700 if (entity->attr.parameter.doubleword_low_mode != NULL) {
701 ir_mode *mode = entity->attr.parameter.doubleword_low_mode;
702 ir_node *val0 = new_r_Proj(args, mode, arg);
703 ir_node *val1 = new_r_Proj(args, mode, arg+1);
704 ir_node *store0 = new_r_Store(start_block, mem, addr, val0,
706 ir_node *mem0 = new_r_Proj(store0, mode_M, pn_Store_M);
707 size_t offset = get_mode_size_bits(mode)/8;
708 ir_mode *addr_mode = get_irn_mode(addr);
709 ir_node *cnst = new_r_Const_long(irg, addr_mode, offset);
710 ir_node *next_addr = new_r_Add(start_block, addr, cnst, addr_mode);
711 ir_node *store1 = new_r_Store(start_block, mem0, next_addr, val1,
713 mem = new_r_Proj(store1, mode_M, pn_Store_M);
714 if (first_store == NULL)
715 first_store = store0;
717 ir_type *tp = get_entity_type(entity);
718 ir_mode *mode = is_compound_type(tp) ? mode_P : get_type_mode(tp);
719 ir_node *val = new_r_Proj(args, mode, arg);
720 ir_node *store = new_r_Store(start_block, mem, addr, val, cons_none);
721 mem = new_r_Proj(store, mode_M, pn_Store_M);
722 if (first_store == NULL)
727 if (mem != initial_mem) {
728 edges_reroute(initial_mem, mem);
729 set_Store_mem(first_store, initial_mem);
733 void be_add_parameter_entity_stores(ir_graph *irg)
735 ir_type *frame_type = get_irg_frame_type(irg);
736 be_stack_layout_t *layout = be_get_irg_stack_layout(irg);
737 ir_type *between_type = layout->between_type;
739 create_stores_for_type(irg, frame_type);
740 if (between_type != NULL) {
741 create_stores_for_type(irg, between_type);