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"
36 #include "irnodemap.h"
41 * An entry in the register state map.
43 typedef struct reg_flag_t {
44 const arch_register_t *reg; /**< register at an input position.
45 may be NULL in case of memory input */
46 arch_register_req_type_t flags; /**< requirement flags for this register. */
50 * A register state mapping keeps track of the symbol values (=firm nodes)
51 * to registers. This is useful when constructing straight line code
52 * like the function prolog or epilog in some architectures.
54 typedef struct register_state_mapping_t {
55 ir_node **value_map; /**< mapping of state indices to values */
56 size_t **reg_index_map; /**< mapping of regclass,regnum to an index
58 reg_flag_t *regs; /**< registers (and memory values) that form a
60 } register_state_mapping_t;
63 * The environment for all helper functions.
65 struct beabi_helper_env_t {
66 ir_graph *irg; /**< the graph we operate on */
67 register_state_mapping_t prolog; /**< the register state map for the prolog */
68 register_state_mapping_t epilog; /**< the register state map for the epilog */
72 * Create a new empty register state map for the given
75 * @param rsm the register state map to be initialized
76 * @param arch_env the architecture environment
78 * After this call, the register map is initialized to empty.
80 static void prepare_rsm(register_state_mapping_t *rsm,
81 const arch_env_t *arch_env)
83 unsigned n_reg_classes = arch_env->n_register_classes;
85 reg_flag_t memory = { NULL, arch_register_req_type_none };
87 rsm->regs = NEW_ARR_F(reg_flag_t, 0);
88 /* memory input at 0 */
89 ARR_APP1(reg_flag_t, rsm->regs, memory);
91 rsm->value_map = NULL;
92 rsm->reg_index_map = XMALLOCN(size_t*, n_reg_classes);
93 for (c = 0; c < n_reg_classes; ++c) {
94 const arch_register_class_t *cls = &arch_env->register_classes[c];
95 unsigned n_regs = arch_register_class_n_regs(cls);
98 rsm->reg_index_map[c] = XMALLOCN(size_t, n_regs);
99 for (r = 0; r < n_regs; ++r) {
100 rsm->reg_index_map[c][r] = (size_t)-1;
106 * Destroy a register state map for the given
109 * @param rsm the register state map to be destroyed
110 * @param arch_env the architecture environment
112 * After this call, the register map is initialized to empty.
114 static void free_rsm(register_state_mapping_t *rsm, const arch_env_t *arch_env)
116 unsigned n_reg_classes = arch_env->n_register_classes;
119 for (c = 0; c < n_reg_classes; ++c) {
120 free(rsm->reg_index_map[c]);
123 free(rsm->reg_index_map);
124 if (rsm->value_map != NULL)
125 DEL_ARR_F(rsm->value_map);
126 DEL_ARR_F(rsm->regs);
129 rsm->reg_index_map = NULL;
130 rsm->value_map = NULL;
134 * Remove all registers from a register state map.
136 * @param rsm the register state map to be destroyed
137 * @param arch_env the architecture environment
139 static void rsm_clear_regs(register_state_mapping_t *rsm,
140 const arch_env_t *arch_env)
142 unsigned n_reg_classes = arch_env->n_register_classes;
144 reg_flag_t memory = { NULL, arch_register_req_type_none };
146 for (c = 0; c < n_reg_classes; ++c) {
147 const arch_register_class_t *cls = &arch_env->register_classes[c];
148 unsigned n_regs = arch_register_class_n_regs(cls);
151 for (r = 0; r < n_regs; ++r) {
152 rsm->reg_index_map[c][r] = (size_t)-1;
155 ARR_RESIZE(reg_flag_t, rsm->regs, 0);
156 ARR_APP1(reg_flag_t, rsm->regs, memory);
158 if (rsm->value_map != NULL) {
159 DEL_ARR_F(rsm->value_map);
160 rsm->value_map = NULL;
165 * Add a register and its constraint flags to a register state map
166 * and return its index inside the map.
168 static size_t rsm_add_reg(register_state_mapping_t *rsm,
169 const arch_register_t *reg,
170 arch_register_req_type_t flags)
172 size_t input_idx = ARR_LEN(rsm->regs);
173 int cls_idx = reg->reg_class->index;
174 int reg_idx = reg->index;
175 reg_flag_t regflag = { reg, flags };
177 /* we must not have used get_value yet */
178 assert(rsm->reg_index_map[cls_idx][reg_idx] == (size_t)-1);
179 rsm->reg_index_map[cls_idx][reg_idx] = input_idx;
180 ARR_APP1(reg_flag_t, rsm->regs, regflag);
182 if (rsm->value_map != NULL) {
183 ARR_APP1(ir_node*, rsm->value_map, NULL);
184 assert(ARR_LEN(rsm->value_map) == ARR_LEN(rsm->regs));
190 * Retrieve the ir_node stored at the given index in the register state map.
192 static ir_node *rsm_get_value(register_state_mapping_t *rsm, size_t index)
194 assert(index < ARR_LEN(rsm->value_map));
195 return rsm->value_map[index];
199 * Retrieve the ir_node occupying the given register in the register state map.
201 static ir_node *rsm_get_reg_value(register_state_mapping_t *rsm,
202 const arch_register_t *reg)
204 int cls_idx = reg->reg_class->index;
205 int reg_idx = reg->index;
206 size_t input_idx = rsm->reg_index_map[cls_idx][reg_idx];
208 return rsm_get_value(rsm, input_idx);
212 * Enter a ir_node at the given index in the register state map.
214 static void rsm_set_value(register_state_mapping_t *rsm, size_t index,
217 assert(index < ARR_LEN(rsm->value_map));
218 rsm->value_map[index] = value;
222 * Enter a ir_node at the given register in the register state map.
224 static void rsm_set_reg_value(register_state_mapping_t *rsm,
225 const arch_register_t *reg, ir_node *value)
227 int cls_idx = reg->reg_class->index;
228 int reg_idx = reg->index;
229 size_t input_idx = rsm->reg_index_map[cls_idx][reg_idx];
230 rsm_set_value(rsm, input_idx, value);
234 beabi_helper_env_t *be_abihelper_prepare(ir_graph *irg)
236 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
237 beabi_helper_env_t *env = XMALLOCZ(beabi_helper_env_t);
240 prepare_rsm(&env->prolog, arch_env);
241 prepare_rsm(&env->epilog, arch_env);
246 void be_abihelper_finish(beabi_helper_env_t *env)
248 const arch_env_t *arch_env = be_get_irg_arch_env(env->irg);
250 free_rsm(&env->prolog, arch_env);
251 if (env->epilog.reg_index_map != NULL) {
252 free_rsm(&env->epilog, arch_env);
257 void be_prolog_add_reg(beabi_helper_env_t *env, const arch_register_t *reg,
258 arch_register_req_type_t flags)
260 rsm_add_reg(&env->prolog, reg, flags);
263 ir_node *be_prolog_create_start(beabi_helper_env_t *env, dbg_info *dbgi,
266 int n_start_outs = ARR_LEN(env->prolog.regs);
267 ir_node *start = be_new_Start(dbgi, block, n_start_outs);
270 assert(env->prolog.value_map == NULL);
271 env->prolog.value_map = NEW_ARR_F(ir_node*, n_start_outs);
273 for (o = 0; o < n_start_outs; ++o) {
274 const reg_flag_t *regflag = &env->prolog.regs[o];
275 const arch_register_t *reg = regflag->reg;
278 arch_set_irn_register_req_out(start, o, arch_no_register_req);
279 proj = new_r_Proj(start, mode_M, o);
281 be_set_constr_single_reg_out(start, o, regflag->reg,
283 arch_set_irn_register_out(start, o, regflag->reg);
284 proj = new_r_Proj(start, reg->reg_class->mode, o);
286 env->prolog.value_map[o] = proj;
292 ir_node *be_prolog_get_reg_value(beabi_helper_env_t *env,
293 const arch_register_t *reg)
295 return rsm_get_reg_value(&env->prolog, reg);
298 ir_node *be_prolog_get_memory(beabi_helper_env_t *env)
300 return rsm_get_value(&env->prolog, 0);
303 void be_prolog_set_reg_value(beabi_helper_env_t *env,
304 const arch_register_t *reg, ir_node *value)
306 rsm_set_reg_value(&env->prolog, reg, value);
309 void be_prolog_set_memory(beabi_helper_env_t *env, ir_node *value)
311 rsm_set_value(&env->prolog, 0, value);
316 void be_epilog_begin(beabi_helper_env_t *env)
318 const arch_env_t *arch_env = be_get_irg_arch_env(env->irg);
319 rsm_clear_regs(&env->epilog, arch_env);
320 env->epilog.value_map = NEW_ARR_F(ir_node*, 1);
321 env->epilog.value_map[0] = NULL;
324 void be_epilog_add_reg(beabi_helper_env_t *env, const arch_register_t *reg,
325 arch_register_req_type_t flags, ir_node *value)
327 size_t index = rsm_add_reg(&env->epilog, reg, flags);
328 rsm_set_value(&env->epilog, index, value);
331 void be_epilog_set_reg_value(beabi_helper_env_t *env,
332 const arch_register_t *reg, ir_node *value)
334 rsm_set_reg_value(&env->epilog, reg, value);
337 void be_epilog_set_memory(beabi_helper_env_t *env, ir_node *value)
339 rsm_set_value(&env->epilog, 0, value);
342 ir_node *be_epilog_get_reg_value(beabi_helper_env_t *env,
343 const arch_register_t *reg)
345 return rsm_get_reg_value(&env->epilog, reg);
348 ir_node *be_epilog_get_memory(beabi_helper_env_t *env)
350 return rsm_get_value(&env->epilog, 0);
353 ir_node *be_epilog_create_return(beabi_helper_env_t *env, dbg_info *dbgi,
356 size_t n_return_in = ARR_LEN(env->epilog.regs);
357 ir_node **in = env->epilog.value_map;
358 int n_res = 1; /* TODO */
359 unsigned pop = 0; /* TODO */
363 assert(ARR_LEN(env->epilog.value_map) == n_return_in);
365 ret = be_new_Return(dbgi, get_irn_irg(block), block, n_res, pop,
367 for (i = 0; i < n_return_in; ++i) {
368 const reg_flag_t *regflag = &env->epilog.regs[i];
369 const arch_register_t *reg = regflag->reg;
371 be_set_constr_single_reg_in(ret, i, reg,
372 arch_register_req_type_none);
376 rsm_clear_regs(&env->epilog, be_get_irg_arch_env(env->irg));
382 * Tests whether a node has a real user and is not just kept by the End or
385 static bool has_real_user(const ir_node *node)
387 foreach_out_edge(node, edge) {
388 ir_node *user = get_edge_src_irn(edge);
389 if (!is_End(user) && !is_Anchor(user))
395 static ir_node *add_to_keep(ir_node *last_keep,
396 const arch_register_class_t *cls, ir_node *node)
398 if (last_keep != NULL) {
399 be_Keep_add_node(last_keep, cls, node);
401 ir_node *in[1] = { node };
402 ir_node *block = get_nodes_block(node);
404 last_keep = be_new_Keep(block, 1, in);
406 schedpoint = skip_Proj(node);
407 if (sched_is_scheduled(schedpoint)) {
408 sched_add_after(schedpoint, last_keep);
414 void be_add_missing_keeps_node(ir_node *node)
417 unsigned *found_projs;
418 ir_mode *mode = get_irn_mode(node);
420 ir_node **existing_projs;
422 if (mode != mode_T) {
423 if (!has_real_user(node)) {
424 const arch_register_req_t *req = arch_get_irn_register_req(node);
425 const arch_register_class_t *cls = req->cls;
427 || (cls->flags & arch_register_class_flag_manual_ra)) {
431 add_to_keep(NULL, cls, node);
436 n_outs = arch_get_irn_n_outs(node);
440 rbitset_alloca(found_projs, n_outs);
441 existing_projs = ALLOCANZ(ir_node*, n_outs);
442 foreach_out_edge(node, edge) {
443 ir_node *succ = get_edge_src_irn(edge);
444 ir_mode *mode = get_irn_mode(succ);
447 /* The node could be kept */
448 if (is_End(succ) || is_Anchor(succ))
450 if (mode == mode_M || mode == mode_X)
452 pn = get_Proj_proj(succ);
453 existing_projs[pn] = succ;
454 if (!has_real_user(succ))
458 rbitset_set(found_projs, pn);
461 /* are keeps missing? */
463 for (i = 0; i < n_outs; ++i) {
465 const arch_register_req_t *req;
466 const arch_register_class_t *cls;
468 if (rbitset_is_set(found_projs, i)) {
472 req = arch_get_irn_register_req_out(node, i);
474 if (cls == NULL || (cls->flags & arch_register_class_flag_manual_ra)) {
478 value = existing_projs[i];
480 value = new_r_Proj(node, arch_register_class_mode(cls), i);
481 last_keep = add_to_keep(last_keep, cls, value);
485 static void add_missing_keep_walker(ir_node *node, void *data)
488 be_add_missing_keeps_node(node);
491 void be_add_missing_keeps(ir_graph *irg)
493 irg_walk_graph(irg, add_missing_keep_walker, NULL, NULL);
498 * Link the node into its block list as a new head.
500 static void collect_node(ir_node *node)
502 ir_node *block = get_nodes_block(node);
503 ir_node *old = (ir_node*)get_irn_link(block);
505 set_irn_link(node, old);
506 set_irn_link(block, node);
510 * Post-walker: link all nodes that probably access the stack into lists of their block.
512 static void link_ops_in_block_walker(ir_node *node, void *data)
516 switch (get_irn_opcode(node)) {
522 /** all non-stack alloc nodes should be lowered before the backend */
523 assert(get_Alloc_where(node) == stack_alloc);
527 assert(get_Free_where(node) == stack_alloc);
531 if (get_Builtin_kind(node) == ir_bk_return_address) {
532 ir_node *param = get_Builtin_param(node, 0);
533 ir_tarval *tv = get_Const_tarval(param); /* must be Const */
534 long value = get_tarval_long(tv);
536 /* not the return address of the current function:
537 * we need the stack pointer for the frame climbing */
547 static ir_heights_t *heights;
550 * Check if a node is somehow data dependent on another one.
551 * both nodes must be in the same basic block.
552 * @param n1 The first node.
553 * @param n2 The second node.
554 * @return 1, if n1 is data dependent (transitively) on n2, 0 if not.
556 static int dependent_on(const ir_node *n1, const ir_node *n2)
558 assert(get_nodes_block(n1) == get_nodes_block(n2));
560 return heights_reachable_in_block(heights, n1, n2);
564 * Classical qsort() comparison function behavior:
566 * 0 if both elements are equal, no node depend on the other
567 * +1 if first depends on second (first is greater)
568 * -1 if second depends on first (second is greater)
570 static int cmp_call_dependency(const void *c1, const void *c2)
572 const ir_node *n1 = *(const ir_node **) c1;
573 const ir_node *n2 = *(const ir_node **) c2;
576 if (dependent_on(n1, n2))
579 if (dependent_on(n2, n1))
582 /* The nodes have no depth order, but we need a total order because qsort()
585 * Additionally, we need to respect transitive dependencies. Consider a
586 * Call a depending on Call b and an independent Call c.
587 * We MUST NOT order c > a and b > c. */
588 h1 = get_irn_height(heights, n1);
589 h2 = get_irn_height(heights, n2);
590 if (h1 < h2) return 1;
591 if (h1 > h2) return -1;
592 /* Same height, so use a random (but stable) order */
593 return get_irn_idx(n2) - get_irn_idx(n1);
597 * Block-walker: sorts dependencies and remember them into a phase
599 static void process_ops_in_block(ir_node *block, void *data)
601 ir_nodemap *map = (ir_nodemap*)data;
608 for (node = (ir_node*)get_irn_link(block); node != NULL;
609 node = (ir_node*)get_irn_link(node)) {
616 nodes = XMALLOCN(ir_node*, n_nodes);
618 for (node = (ir_node*)get_irn_link(block); node != NULL;
619 node = (ir_node*)get_irn_link(node)) {
622 assert(n == n_nodes);
624 /* order nodes according to their data dependencies */
625 qsort(nodes, n_nodes, sizeof(nodes[0]), cmp_call_dependency);
627 /* remember the calculated dependency into a phase */
628 for (n = n_nodes-1; n > 0; --n) {
629 ir_node *node = nodes[n];
630 ir_node *pred = nodes[n-1];
632 ir_nodemap_insert(map, node, pred);
639 struct be_stackorder_t {
640 ir_nodemap stack_order; /**< a phase to handle stack dependencies. */
643 be_stackorder_t *be_collect_stacknodes(ir_graph *irg)
645 be_stackorder_t *env = XMALLOCZ(be_stackorder_t);
647 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
649 /* collect all potential^stack accessing nodes */
650 irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, NULL);
652 ir_nodemap_init(&env->stack_order, irg);
654 /* use heights to create a total order for those nodes: this order is stored
655 * in the created phase */
656 heights = heights_new(irg);
657 irg_block_walk_graph(irg, NULL, process_ops_in_block, &env->stack_order);
658 heights_free(heights);
660 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
665 ir_node *be_get_stack_pred(const be_stackorder_t *env, const ir_node *node)
667 return ir_nodemap_get(ir_node, &env->stack_order, node);
670 void be_free_stackorder(be_stackorder_t *env)
672 ir_nodemap_destroy(&env->stack_order);
676 static void create_stores_for_type(ir_graph *irg, ir_type *type)
678 size_t n = get_compound_n_members(type);
679 ir_node *frame = get_irg_frame(irg);
680 ir_node *initial_mem = get_irg_initial_mem(irg);
681 ir_node *mem = initial_mem;
682 ir_node *first_store = NULL;
683 ir_node *start_block = get_irg_start_block(irg);
684 ir_node *args = get_irg_args(irg);
687 /* all parameter entities left in the frame type require stores.
688 * (The ones passed on the stack have been moved to the arg type) */
689 for (i = 0; i < n; ++i) {
690 ir_entity *entity = get_compound_member(type, i);
693 if (!is_parameter_entity(entity))
696 arg = get_entity_parameter_number(entity);
697 if (arg == IR_VA_START_PARAMETER_NUMBER)
700 addr = new_r_Sel(start_block, mem, frame, 0, NULL, entity);
701 if (entity->attr.parameter.doubleword_low_mode != NULL) {
702 ir_mode *mode = entity->attr.parameter.doubleword_low_mode;
703 ir_node *val0 = new_r_Proj(args, mode, arg);
704 ir_node *val1 = new_r_Proj(args, mode, arg+1);
705 ir_node *store0 = new_r_Store(start_block, mem, addr, val0,
707 ir_node *mem0 = new_r_Proj(store0, mode_M, pn_Store_M);
708 size_t offset = get_mode_size_bits(mode)/8;
709 ir_mode *addr_mode = get_irn_mode(addr);
710 ir_node *cnst = new_r_Const_long(irg, addr_mode, offset);
711 ir_node *next_addr = new_r_Add(start_block, addr, cnst, addr_mode);
712 ir_node *store1 = new_r_Store(start_block, mem0, next_addr, val1,
714 mem = new_r_Proj(store1, mode_M, pn_Store_M);
715 if (first_store == NULL)
716 first_store = store0;
718 ir_type *tp = get_entity_type(entity);
719 ir_mode *mode = is_compound_type(tp) ? mode_P : get_type_mode(tp);
720 ir_node *val = new_r_Proj(args, mode, arg);
721 ir_node *store = new_r_Store(start_block, mem, addr, val, cons_none);
722 mem = new_r_Proj(store, mode_M, pn_Store_M);
723 if (first_store == NULL)
728 if (mem != initial_mem)
729 edges_reroute_except(initial_mem, mem, first_store);
732 void be_add_parameter_entity_stores(ir_graph *irg)
734 ir_type *frame_type = get_irg_frame_type(irg);
735 be_stack_layout_t *layout = be_get_irg_stack_layout(irg);
736 ir_type *between_type = layout->between_type;
738 create_stores_for_type(irg, frame_type);
739 if (between_type != NULL) {
740 create_stores_for_type(irg, between_type);