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"
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 */
68 ir_phase *stack_order; /**< a phase to handle stack dependencies. */
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 int 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 arch_irn_add_flags(start, arch_irn_flags_prolog);
272 assert(env->prolog.value_map == NULL);
273 env->prolog.value_map = NEW_ARR_F(ir_node*, n_start_outs);
275 for (o = 0; o < n_start_outs; ++o) {
276 const reg_flag_t *regflag = &env->prolog.regs[o];
277 const arch_register_t *reg = regflag->reg;
280 arch_set_out_register_req(start, o, arch_no_register_req);
281 proj = new_r_Proj(start, mode_M, o);
283 be_set_constr_single_reg_out(start, o, regflag->reg,
285 arch_irn_set_register(start, o, regflag->reg);
286 proj = new_r_Proj(start, reg->reg_class->mode, o);
288 env->prolog.value_map[o] = proj;
294 ir_node *be_prolog_get_reg_value(beabi_helper_env_t *env,
295 const arch_register_t *reg)
297 return rsm_get_reg_value(&env->prolog, reg);
300 ir_node *be_prolog_get_memory(beabi_helper_env_t *env)
302 return rsm_get_value(&env->prolog, 0);
305 void be_prolog_set_reg_value(beabi_helper_env_t *env,
306 const arch_register_t *reg, ir_node *value)
308 rsm_set_reg_value(&env->prolog, reg, value);
311 void be_prolog_set_memory(beabi_helper_env_t *env, ir_node *value)
313 rsm_set_value(&env->prolog, 0, value);
318 void be_epilog_begin(beabi_helper_env_t *env)
320 const arch_env_t *arch_env = be_get_irg_arch_env(env->irg);
321 rsm_clear_regs(&env->epilog, arch_env);
322 env->epilog.value_map = NEW_ARR_F(ir_node*, 1);
323 env->epilog.value_map[0] = NULL;
326 void be_epilog_add_reg(beabi_helper_env_t *env, const arch_register_t *reg,
327 arch_register_req_type_t flags, ir_node *value)
329 int index = rsm_add_reg(&env->epilog, reg, flags);
330 rsm_set_value(&env->epilog, index, value);
333 void be_epilog_set_reg_value(beabi_helper_env_t *env,
334 const arch_register_t *reg, ir_node *value)
336 rsm_set_reg_value(&env->epilog, reg, value);
339 void be_epilog_set_memory(beabi_helper_env_t *env, ir_node *value)
341 rsm_set_value(&env->epilog, 0, value);
344 ir_node *be_epilog_get_reg_value(beabi_helper_env_t *env,
345 const arch_register_t *reg)
347 return rsm_get_reg_value(&env->epilog, reg);
350 ir_node *be_epilog_get_memory(beabi_helper_env_t *env)
352 return rsm_get_value(&env->epilog, 0);
355 ir_node *be_epilog_create_return(beabi_helper_env_t *env, dbg_info *dbgi,
358 size_t n_return_in = ARR_LEN(env->epilog.regs);
359 ir_node **in = env->epilog.value_map;
360 int n_res = 1; /* TODO */
361 unsigned pop = 0; /* TODO */
365 assert(ARR_LEN(env->epilog.value_map) == n_return_in);
367 ret = be_new_Return(dbgi, get_irn_irg(block), block, n_res, pop,
369 arch_irn_add_flags(ret, arch_irn_flags_epilog);
370 for (i = 0; i < n_return_in; ++i) {
371 const reg_flag_t *regflag = &env->epilog.regs[i];
372 const arch_register_t *reg = regflag->reg;
374 be_set_constr_single_reg_in(ret, i, reg,
375 arch_register_req_type_none);
379 rsm_clear_regs(&env->epilog, be_get_irg_arch_env(env->irg));
385 * Tests whether a node has a real user and is not just kept by the End or
388 static bool has_real_user(const ir_node *node)
390 const ir_edge_t *edge;
391 foreach_out_edge(node, edge) {
392 ir_node *user = get_edge_src_irn(edge);
393 if (!is_End(user) && !is_Anchor(user))
399 static ir_node *add_to_keep(ir_node *last_keep,
400 const arch_register_class_t *cls, ir_node *node)
403 if (last_keep != NULL) {
404 be_Keep_add_node(last_keep, cls, node);
406 ir_node *in[1] = { node };
407 ir_node *block = get_nodes_block(node);
409 last_keep = be_new_Keep(block, 1, in);
411 schedpoint = skip_Proj(node);
412 if (sched_is_scheduled(schedpoint)) {
413 sched_add_after(schedpoint, last_keep);
416 op = skip_Proj_const(node);
417 if (arch_irn_get_flags(op) & arch_irn_flags_prolog)
418 arch_irn_add_flags(last_keep, arch_irn_flags_prolog);
419 if (arch_irn_get_flags(op) & arch_irn_flags_epilog)
420 arch_irn_add_flags(last_keep, arch_irn_flags_epilog);
424 static void add_missing_keep_walker(ir_node *node, void *data)
427 unsigned *found_projs;
428 const ir_edge_t *edge;
429 ir_mode *mode = get_irn_mode(node);
432 if (mode != mode_T) {
433 if (!has_real_user(node)) {
434 const arch_register_req_t *req = arch_get_register_req_out(node);
435 const arch_register_class_t *cls = req->cls;
437 || (cls->flags & arch_register_class_flag_manual_ra)) {
441 add_to_keep(NULL, cls, node);
446 n_outs = arch_irn_get_n_outs(node);
450 rbitset_alloca(found_projs, n_outs);
451 foreach_out_edge(node, edge) {
452 ir_node *succ = get_edge_src_irn(edge);
453 ir_mode *mode = get_irn_mode(succ);
456 /* The node could be kept */
457 if (is_End(succ) || is_Anchor(succ))
459 if (mode == mode_M || mode == mode_X)
461 if (!has_real_user(succ))
464 pn = get_Proj_proj(succ);
466 rbitset_set(found_projs, pn);
469 /* are keeps missing? */
471 for (i = 0; i < n_outs; ++i) {
473 const arch_register_req_t *req;
474 const arch_register_class_t *cls;
476 if (rbitset_is_set(found_projs, i)) {
480 req = arch_get_out_register_req(node, i);
482 if (cls == NULL || (cls->flags & arch_register_class_flag_manual_ra)) {
486 value = new_r_Proj(node, arch_register_class_mode(cls), i);
487 last_keep = add_to_keep(last_keep, cls, value);
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;
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()
583 return get_irn_idx(n2) - get_irn_idx(n1);
587 * Block-walker: sorts dependencies and remember them into a phase
589 static void process_ops_in_block(ir_node *block, void *data)
591 ir_phase *phase = (ir_phase*)data;
598 for (node = (ir_node*)get_irn_link(block); node != NULL;
599 node = (ir_node*)get_irn_link(node)) {
606 nodes = XMALLOCN(ir_node*, n_nodes);
608 for (node = (ir_node*)get_irn_link(block); node != NULL;
609 node = (ir_node*)get_irn_link(node)) {
612 assert(n == n_nodes);
614 /* order nodes according to their data dependencies */
615 qsort(nodes, n_nodes, sizeof(nodes[0]), cmp_call_dependency);
617 /* remember the calculated dependency into a phase */
618 for (n = n_nodes-1; n > 0; --n) {
619 ir_node *node = nodes[n];
620 ir_node *pred = nodes[n-1];
622 phase_set_irn_data(phase, node, pred);
627 void be_collect_stacknodes(beabi_helper_env_t *env)
629 ir_graph *irg = env->irg;
631 /* collect all potential^stack accessing nodes */
632 irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, NULL);
634 assert(env->stack_order == NULL);
635 env->stack_order = new_phase(irg, phase_irn_init_default);
637 /* use heights to create a total order for those nodes: this order is stored
638 * in the created phase */
639 heights = heights_new(irg);
640 irg_block_walk_graph(irg, NULL, process_ops_in_block, env->stack_order);
641 heights_free(heights);
644 ir_node *be_get_stack_pred(const beabi_helper_env_t *env, const ir_node *node)
646 return (ir_node*)phase_get_irn_data(env->stack_order, node);