2 * Copyright (C) 1995-2011 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 Backend ABI implementation.
23 * @author Sebastian Hack, Michael Beck
32 #include "irgraph_t.h"
35 #include "iredges_t.h"
38 #include "irprintf_t.h"
45 #include "raw_bitset.h"
56 #include "bessaconstr.h"
58 #include "betranshlp.h"
60 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
62 typedef struct be_abi_call_arg_t {
63 unsigned is_res : 1; /**< 1: the call argument is a return value. 0: it's a call parameter. */
64 unsigned in_reg : 1; /**< 1: this argument is transmitted in registers. */
65 unsigned on_stack : 1; /**< 1: this argument is transmitted on the stack. */
66 unsigned callee : 1; /**< 1: someone called us. 0: We call another function */
69 const arch_register_t *reg;
72 unsigned alignment; /**< stack alignment */
73 unsigned space_before; /**< allocate space before */
74 unsigned space_after; /**< allocate space after */
77 struct be_abi_call_t {
78 be_abi_call_flags_t flags; /**< Flags describing the ABI behavior on calls */
79 int pop; /**< number of bytes the stack frame is shrinked by the callee on return. */
80 const be_abi_callbacks_t *cb;
81 ir_type *between_type;
83 const arch_register_class_t *cls_addr; /**< register class of the call address */
87 * The ABI information for the current graph.
90 be_abi_call_t *call; /**< The ABI call information. */
92 ir_node *init_sp; /**< The node representing the stack pointer
93 at the start of the function. */
95 ir_node *start; /**< The be_Start params node. */
96 pmap *regs; /**< A map of all callee-save and ignore regs to
97 their Projs to the RegParams node. */
98 int start_block_bias; /**< The stack bias at the end of the start block. */
100 pmap *keep_map; /**< mapping blocks to keep nodes. */
102 ir_node **calls; /**< flexible array containing all be_Call nodes */
105 static ir_heights_t *ir_heights;
107 /** Flag: if set, try to omit the frame pointer in all routines. */
108 static int be_omit_fp = 1;
110 static ir_node *be_abi_reg_map_get(pmap *map, const arch_register_t *reg)
112 return (ir_node*)pmap_get(map, reg);
115 static void be_abi_reg_map_set(pmap *map, const arch_register_t* reg,
118 pmap_insert(map, reg, node);
122 _ ____ ___ ____ _ _ _ _
123 / \ | __ )_ _| / ___|__ _| | | |__ __ _ ___| | _____
124 / _ \ | _ \| | | | / _` | | | '_ \ / _` |/ __| |/ / __|
125 / ___ \| |_) | | | |__| (_| | | | |_) | (_| | (__| <\__ \
126 /_/ \_\____/___| \____\__,_|_|_|_.__/ \__,_|\___|_|\_\___/
128 These callbacks are used by the backend to set the parameters
129 for a specific call type.
133 * Set compare function: compares two ABI call object arguments.
135 static int cmp_call_arg(const void *a, const void *b, size_t n)
137 const be_abi_call_arg_t *p = (const be_abi_call_arg_t*)a;
138 const be_abi_call_arg_t *q = (const be_abi_call_arg_t*)b;
140 return !(p->is_res == q->is_res && p->pos == q->pos && p->callee == q->callee);
144 * Get an ABI call object argument.
146 * @param call the abi call
147 * @param is_res true for call results, false for call arguments
148 * @param pos position of the argument
149 * @param callee context type - if we are callee or caller
151 static be_abi_call_arg_t *get_call_arg(be_abi_call_t *call, int is_res, int pos, int callee)
153 be_abi_call_arg_t arg;
156 memset(&arg, 0, sizeof(arg));
161 hash = is_res * 128 + pos;
163 return (be_abi_call_arg_t*)set_find(call->params, &arg, sizeof(arg), hash);
167 * Set an ABI call object argument.
169 static void remember_call_arg(be_abi_call_arg_t *arg, be_abi_call_t *call, be_abi_context_t context)
171 unsigned hash = arg->is_res * 128 + arg->pos;
172 if (context & ABI_CONTEXT_CALLEE) {
174 set_insert(call->params, arg, sizeof(*arg), hash);
176 if (context & ABI_CONTEXT_CALLER) {
178 set_insert(call->params, arg, sizeof(*arg), hash);
182 /* Set the flags for a call. */
183 void be_abi_call_set_flags(be_abi_call_t *call, be_abi_call_flags_t flags, const be_abi_callbacks_t *cb)
189 /* Sets the number of bytes the stackframe is shrinked by the callee on return */
190 void be_abi_call_set_pop(be_abi_call_t *call, int pop)
196 /* Set register class for call address */
197 void be_abi_call_set_call_address_reg_class(be_abi_call_t *call, const arch_register_class_t *cls)
199 call->cls_addr = cls;
203 void be_abi_call_param_stack(be_abi_call_t *call, int arg_pos,
204 ir_mode *load_mode, unsigned alignment,
205 unsigned space_before, unsigned space_after,
206 be_abi_context_t context)
208 be_abi_call_arg_t arg;
209 memset(&arg, 0, sizeof(arg));
210 assert(alignment > 0 && "Alignment must be greater than 0");
212 arg.load_mode = load_mode;
213 arg.alignment = alignment;
214 arg.space_before = space_before;
215 arg.space_after = space_after;
219 remember_call_arg(&arg, call, context);
222 void be_abi_call_param_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
224 be_abi_call_arg_t arg;
225 memset(&arg, 0, sizeof(arg));
232 remember_call_arg(&arg, call, context);
235 void be_abi_call_res_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
237 be_abi_call_arg_t arg;
238 memset(&arg, 0, sizeof(arg));
245 remember_call_arg(&arg, call, context);
248 /* Get the flags of a ABI call object. */
249 be_abi_call_flags_t be_abi_call_get_flags(const be_abi_call_t *call)
255 * Constructor for a new ABI call object.
257 * @param cls_addr register class of the call address
259 * @return the new ABI call object
261 static be_abi_call_t *be_abi_call_new(const arch_register_class_t *cls_addr)
263 be_abi_call_t *call = XMALLOCZ(be_abi_call_t);
266 call->params = new_set(cmp_call_arg, 16);
268 call->cls_addr = cls_addr;
270 call->flags.bits.try_omit_fp = be_omit_fp;
276 * Destructor for an ABI call object.
278 static void be_abi_call_free(be_abi_call_t *call)
280 del_set(call->params);
285 * Initializes the frame layout from parts
287 * @param frame the stack layout that will be initialized
288 * @param args the stack argument layout type
289 * @param between the between layout type
290 * @param locals the method frame type
291 * @param param_map an array mapping method argument positions to the stack argument type
293 * @return the initialized stack layout
295 static be_stack_layout_t *stack_frame_init(be_stack_layout_t *frame, ir_type *args,
296 ir_type *between, ir_type *locals,
297 ir_entity *param_map[])
299 frame->arg_type = args;
300 frame->between_type = between;
301 frame->frame_type = locals;
302 frame->initial_offset = 0;
303 frame->initial_bias = 0;
304 frame->order[1] = between;
305 frame->param_map = param_map;
307 /* typical decreasing stack: locals have the
308 * lowest addresses, arguments the highest */
309 frame->order[0] = locals;
310 frame->order[2] = args;
321 Adjustment of the calls inside a graph.
326 * Transform a call node into a be_Call node.
328 * @param env The ABI environment for the current irg.
329 * @param irn The call node.
330 * @param curr_sp The stack pointer node to use.
331 * @return The stack pointer after the call.
333 static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
335 ir_graph *irg = get_irn_irg(irn);
336 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
337 ir_type *call_tp = get_Call_type(irn);
338 ir_node *call_ptr = get_Call_ptr(irn);
339 size_t n_params = get_method_n_params(call_tp);
340 ir_node *curr_mem = get_Call_mem(irn);
341 ir_node *bl = get_nodes_block(irn);
343 const arch_register_t *sp = arch_env->sp;
344 be_abi_call_t *call = be_abi_call_new(sp->reg_class);
345 ir_mode *mach_mode = sp->reg_class->mode;
346 int no_alloc = call->flags.bits.frame_is_setup_on_call;
347 int n_res = get_method_n_ress(call_tp);
348 int do_seq = call->flags.bits.store_args_sequential && !no_alloc;
350 ir_node *res_proj = NULL;
351 int n_reg_params = 0;
352 int n_stack_params = 0;
355 const arch_register_t **states = NEW_ARR_F(const arch_register_t*, 0);
356 const arch_register_t **destroyed_regs = NEW_ARR_F(const arch_register_t*, 0);
360 int n_reg_results = 0;
361 const ir_edge_t *edge;
363 int *stack_param_idx;
364 int i, n, destroy_all_regs;
369 /* Let the isa fill out the abi description for that call node. */
370 arch_env_get_call_abi(arch_env, call_tp, call);
372 /* Insert code to put the stack arguments on the stack. */
373 assert(get_Call_n_params(irn) == n_params);
374 stack_param_idx = ALLOCAN(int, n_params);
375 for (p = 0; p < n_params; ++p) {
376 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
379 int arg_size = get_type_size_bytes(get_method_param_type(call_tp, p));
381 stack_size += round_up2(arg->space_before, arg->alignment);
382 stack_size += round_up2(arg_size, arg->alignment);
383 stack_size += round_up2(arg->space_after, arg->alignment);
385 stack_param_idx[n_stack_params++] = p;
389 /* Collect all arguments which are passed in registers. */
390 reg_param_idxs = ALLOCAN(int, n_params);
391 for (p = 0; p < n_params; ++p) {
392 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
393 if (arg && arg->in_reg) {
394 reg_param_idxs[n_reg_params++] = p;
399 * If the stack is decreasing and we do not want to store sequentially,
400 * or someone else allocated the call frame
401 * we allocate as much space on the stack all parameters need, by
402 * moving the stack pointer along the stack's direction.
404 * Note: we also have to do this for stack_size == 0, because we may have
405 * to adjust stack alignment for the call.
407 if (!do_seq && !no_alloc) {
408 curr_sp = be_new_IncSP(sp, bl, curr_sp, stack_size, 1);
411 dbgi = get_irn_dbg_info(irn);
412 /* If there are some parameters which shall be passed on the stack. */
413 if (n_stack_params > 0) {
415 ir_node **in = ALLOCAN(ir_node*, n_stack_params+1);
419 * Reverse list of stack parameters if call arguments are from left to right.
420 * We must them reverse again if they are pushed (not stored) and the stack
421 * direction is downwards.
423 if (call->flags.bits.left_to_right ^ do_seq) {
424 for (i = 0; i < n_stack_params >> 1; ++i) {
425 int other = n_stack_params - i - 1;
426 int tmp = stack_param_idx[i];
427 stack_param_idx[i] = stack_param_idx[other];
428 stack_param_idx[other] = tmp;
432 curr_mem = get_Call_mem(irn);
434 in[n_in++] = curr_mem;
437 for (i = 0; i < n_stack_params; ++i) {
438 int p = stack_param_idx[i];
439 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
440 ir_node *param = get_Call_param(irn, p);
441 ir_node *addr = curr_sp;
443 ir_type *param_type = get_method_param_type(call_tp, p);
444 int param_size = get_type_size_bytes(param_type) + arg->space_after;
447 * If we wanted to build the arguments sequentially,
448 * the stack pointer for the next must be incremented,
449 * and the memory value propagated.
453 addr = curr_sp = be_new_IncSP(sp, bl, curr_sp,
454 param_size + arg->space_before, 0);
455 add_irn_dep(curr_sp, curr_mem);
457 curr_ofs += arg->space_before;
458 curr_ofs = round_up2(curr_ofs, arg->alignment);
460 /* Make the expression to compute the argument's offset. */
462 ir_mode *constmode = mach_mode;
463 if (mode_is_reference(mach_mode)) {
466 addr = new_r_Const_long(irg, constmode, curr_ofs);
467 addr = new_r_Add(bl, curr_sp, addr, mach_mode);
471 /* Insert a store for primitive arguments. */
472 if (is_atomic_type(param_type)) {
473 ir_node *nomem = get_irg_no_mem(irg);
474 ir_node *mem_input = do_seq ? curr_mem : nomem;
475 ir_node *store = new_rd_Store(dbgi, bl, mem_input, addr, param, cons_none);
476 mem = new_r_Proj(store, mode_M, pn_Store_M);
478 /* Make a mem copy for compound arguments. */
481 assert(mode_is_reference(get_irn_mode(param)));
482 copy = new_rd_CopyB(dbgi, bl, curr_mem, addr, param, param_type);
483 mem = new_r_Proj(copy, mode_M, pn_CopyB_M);
486 curr_ofs += param_size;
494 /* We need the sync only, if we didn't build the stores sequentially. */
496 if (n_stack_params >= 1) {
497 curr_mem = new_r_Sync(bl, n_in, in);
499 curr_mem = get_Call_mem(irn);
504 /* check for the return_twice property */
505 destroy_all_regs = 0;
506 if (is_SymConst_addr_ent(call_ptr)) {
507 ir_entity *ent = get_SymConst_entity(call_ptr);
509 if (get_entity_additional_properties(ent) & mtp_property_returns_twice)
510 destroy_all_regs = 1;
512 ir_type *call_tp = get_Call_type(irn);
514 if (get_method_additional_properties(call_tp) & mtp_property_returns_twice)
515 destroy_all_regs = 1;
518 /* Put caller save into the destroyed set and state registers in the states
520 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
522 const arch_register_class_t *cls = &arch_env->register_classes[i];
523 for (j = 0; j < cls->n_regs; ++j) {
524 const arch_register_t *reg = arch_register_for_index(cls, j);
526 /* even if destroyed all is specified, neither SP nor FP are
527 * destroyed (else bad things will happen) */
528 if (reg == arch_env->sp || reg == arch_env->bp)
531 if (reg->type & arch_register_type_state) {
532 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
533 ARR_APP1(const arch_register_t*, states, reg);
534 /* we're already in the destroyed set so no need for further
538 if (destroy_all_regs || (reg->type & arch_register_type_caller_save)) {
539 if (!(reg->type & arch_register_type_ignore)) {
540 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
546 /* search the largest result proj number */
547 res_projs = ALLOCANZ(ir_node*, n_res);
549 foreach_out_edge(irn, edge) {
550 const ir_edge_t *res_edge;
551 ir_node *irn = get_edge_src_irn(edge);
553 if (!is_Proj(irn) || get_Proj_proj(irn) != pn_Call_T_result)
556 foreach_out_edge(irn, res_edge) {
558 ir_node *res = get_edge_src_irn(res_edge);
560 assert(is_Proj(res));
562 proj = get_Proj_proj(res);
563 assert(proj < n_res);
564 assert(res_projs[proj] == NULL);
565 res_projs[proj] = res;
571 /** TODO: this is not correct for cases where return values are passed
572 * on the stack, but no known ABI does this currently...
574 n_reg_results = n_res;
577 in = ALLOCAN(ir_node*, n_reg_params + ARR_LEN(states));
579 /* make the back end call node and set its register requirements. */
580 for (i = 0; i < n_reg_params; ++i) {
581 in[n_ins++] = get_Call_param(irn, reg_param_idxs[i]);
584 /* add state registers ins */
585 for (s = 0; s < ARR_LEN(states); ++s) {
586 const arch_register_t *reg = states[s];
587 const arch_register_class_t *cls = arch_register_get_class(reg);
588 ir_node *regnode = new_r_Unknown(irg, arch_register_class_mode(cls));
589 in[n_ins++] = regnode;
591 assert(n_ins == (int) (n_reg_params + ARR_LEN(states)));
593 /* ins collected, build the call */
594 if (env->call->flags.bits.call_has_imm && is_SymConst(call_ptr)) {
596 low_call = be_new_Call(dbgi, irg, bl, curr_mem, curr_sp, curr_sp,
597 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
598 n_ins, in, get_Call_type(irn));
599 be_Call_set_entity(low_call, get_SymConst_entity(call_ptr));
602 low_call = be_new_Call(dbgi, irg, bl, curr_mem, curr_sp, call_ptr,
603 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
604 n_ins, in, get_Call_type(irn));
606 be_Call_set_pop(low_call, call->pop);
608 /* put the call into the list of all calls for later processing */
609 ARR_APP1(ir_node *, env->calls, low_call);
611 /* create new stack pointer */
612 curr_sp = new_r_Proj(low_call, get_irn_mode(curr_sp), pn_be_Call_sp);
613 be_set_constr_single_reg_out(low_call, pn_be_Call_sp, sp,
614 arch_register_req_type_ignore | arch_register_req_type_produces_sp);
615 arch_set_irn_register(curr_sp, sp);
617 /* now handle results */
618 for (i = 0; i < n_res; ++i) {
620 ir_node *proj = res_projs[i];
621 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
623 /* returns values on stack not supported yet */
627 shift the proj number to the right, since we will drop the
628 unspeakable Proj_T from the Call. Therefore, all real argument
629 Proj numbers must be increased by pn_be_Call_first_res
631 pn = i + pn_be_Call_first_res;
634 ir_type *res_type = get_method_res_type(call_tp, i);
635 ir_mode *mode = get_type_mode(res_type);
636 proj = new_r_Proj(low_call, mode, pn);
639 set_Proj_pred(proj, low_call);
640 set_Proj_proj(proj, pn);
644 /* remove register from destroyed regs */
646 size_t n = ARR_LEN(destroyed_regs);
647 for (j = 0; j < n; ++j) {
648 if (destroyed_regs[j] == arg->reg) {
649 destroyed_regs[j] = destroyed_regs[n-1];
650 ARR_SHRINKLEN(destroyed_regs,n-1);
658 Set the register class of the call address to
659 the backend provided class (default: stack pointer class)
661 be_node_set_reg_class_in(low_call, n_be_Call_ptr, call->cls_addr);
663 DBG((dbg, LEVEL_3, "\tcreated backend call %+F\n", low_call));
665 /* Set the register classes and constraints of the Call parameters. */
666 for (i = 0; i < n_reg_params; ++i) {
667 int index = reg_param_idxs[i];
668 be_abi_call_arg_t *arg = get_call_arg(call, 0, index, 0);
669 assert(arg->reg != NULL);
671 be_set_constr_single_reg_in(low_call, n_be_Call_first_arg + i,
672 arg->reg, arch_register_req_type_none);
675 /* Set the register constraints of the results. */
676 for (i = 0; i < n_res; ++i) {
677 ir_node *proj = res_projs[i];
678 const be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
679 int pn = get_Proj_proj(proj);
682 be_set_constr_single_reg_out(low_call, pn, arg->reg,
683 arch_register_req_type_none);
684 arch_set_irn_register(proj, arg->reg);
686 exchange(irn, low_call);
688 /* kill the ProjT node */
689 if (res_proj != NULL) {
693 /* Make additional projs for the caller save registers
694 and the Keep node which keeps them alive. */
700 int curr_res_proj = pn_be_Call_first_res + n_reg_results;
703 n_ins = ARR_LEN(destroyed_regs) + n_reg_results + 1;
704 in = ALLOCAN(ir_node *, n_ins);
706 /* also keep the stack pointer */
707 set_irn_link(curr_sp, (void*) sp);
710 for (d = 0; d < ARR_LEN(destroyed_regs); ++d) {
711 const arch_register_t *reg = destroyed_regs[d];
712 ir_node *proj = new_r_Proj(low_call, reg->reg_class->mode, curr_res_proj);
714 /* memorize the register in the link field. we need afterwards to set the register class of the keep correctly. */
715 be_set_constr_single_reg_out(low_call, curr_res_proj, reg,
716 arch_register_req_type_none);
717 arch_set_irn_register(proj, reg);
719 set_irn_link(proj, (void*) reg);
724 for (i = 0; i < n_reg_results; ++i) {
725 ir_node *proj = res_projs[i];
726 const arch_register_t *reg = arch_get_irn_register(proj);
727 set_irn_link(proj, (void*) reg);
732 /* create the Keep for the caller save registers */
733 keep = be_new_Keep(bl, n, in);
734 for (i = 0; i < n; ++i) {
735 const arch_register_t *reg = (const arch_register_t*)get_irn_link(in[i]);
736 be_node_set_reg_class_in(keep, i, reg->reg_class);
740 /* Clean up the stack. */
741 assert(stack_size >= call->pop);
742 stack_size -= call->pop;
744 if (stack_size > 0) {
745 ir_node *mem_proj = NULL;
747 foreach_out_edge(low_call, edge) {
748 ir_node *irn = get_edge_src_irn(edge);
749 if (is_Proj(irn) && get_Proj_proj(irn) == pn_Call_M) {
756 mem_proj = new_r_Proj(low_call, mode_M, pn_be_Call_M_regular);
757 keep_alive(mem_proj);
760 /* Clean up the stack frame or revert alignment fixes if we allocated it */
762 curr_sp = be_new_IncSP(sp, bl, curr_sp, -stack_size, 0);
765 be_abi_call_free(call);
768 DEL_ARR_F(destroyed_regs);
774 * Adjust the size of a node representing a stack alloc or free for the minimum stack alignment.
776 * @param alignment the minimum stack alignment
777 * @param size the node containing the non-aligned size
778 * @param block the block where new nodes are allocated on
779 * @param dbg debug info for new nodes
781 * @return a node representing the aligned size
783 static ir_node *adjust_alloc_size(unsigned stack_alignment, ir_node *size,
784 ir_node *block, dbg_info *dbg)
786 if (stack_alignment > 1) {
792 assert(is_po2(stack_alignment));
794 mode = get_irn_mode(size);
795 tv = new_tarval_from_long(stack_alignment-1, mode);
796 irg = get_Block_irg(block);
797 mask = new_r_Const(irg, tv);
798 size = new_rd_Add(dbg, block, size, mask, mode);
800 tv = new_tarval_from_long(-(long)stack_alignment, mode);
801 mask = new_r_Const(irg, tv);
802 size = new_rd_And(dbg, block, size, mask, mode);
808 * The alloca is transformed into a back end alloca node and connected to the stack nodes.
810 static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp)
812 ir_node *block = get_nodes_block(alloc);
813 ir_graph *irg = get_Block_irg(block);
814 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
815 ir_node *alloc_mem = NULL;
816 ir_node *alloc_res = NULL;
817 ir_type *type = get_Alloc_type(alloc);
820 const ir_edge_t *edge;
825 unsigned stack_alignment;
827 /* all non-stack Alloc nodes should already be lowered before the backend */
828 assert(get_Alloc_where(alloc) == stack_alloc);
830 foreach_out_edge(alloc, edge) {
831 ir_node *irn = get_edge_src_irn(edge);
833 assert(is_Proj(irn));
834 switch (get_Proj_proj(irn)) {
846 /* Beware: currently Alloc nodes without a result might happen,
847 only escape analysis kills them and this phase runs only for object
848 oriented source. We kill the Alloc here. */
849 if (alloc_res == NULL && alloc_mem) {
850 exchange(alloc_mem, get_Alloc_mem(alloc));
854 dbg = get_irn_dbg_info(alloc);
855 count = get_Alloc_count(alloc);
857 /* we might need to multiply the count with the element size */
858 if (type != firm_unknown_type && get_type_size_bytes(type) != 1) {
859 ir_mode *mode = get_irn_mode(count);
860 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type),
862 ir_node *cnst = new_rd_Const(dbg, irg, tv);
863 size = new_rd_Mul(dbg, block, count, cnst, mode);
868 /* The stack pointer will be modified in an unknown manner.
869 We cannot omit it. */
870 env->call->flags.bits.try_omit_fp = 0;
872 stack_alignment = 1 << arch_env->stack_alignment;
873 size = adjust_alloc_size(stack_alignment, size, block, dbg);
874 new_alloc = be_new_AddSP(arch_env->sp, block, curr_sp, size);
875 set_irn_dbg_info(new_alloc, dbg);
877 if (alloc_mem != NULL) {
881 addsp_mem = new_r_Proj(new_alloc, mode_M, pn_be_AddSP_M);
883 /* We need to sync the output mem of the AddSP with the input mem
884 edge into the alloc node. */
885 ins[0] = get_Alloc_mem(alloc);
887 sync = new_r_Sync(block, 2, ins);
889 exchange(alloc_mem, sync);
892 exchange(alloc, new_alloc);
894 /* fix projnum of alloca res */
895 set_Proj_proj(alloc_res, pn_be_AddSP_res);
897 curr_sp = new_r_Proj(new_alloc, get_irn_mode(curr_sp), pn_be_AddSP_sp);
904 * The Free is transformed into a back end free node and connected to the stack nodes.
906 static ir_node *adjust_free(be_abi_irg_t *env, ir_node *free, ir_node *curr_sp)
908 ir_node *block = get_nodes_block(free);
909 ir_graph *irg = get_irn_irg(free);
910 ir_type *type = get_Free_type(free);
911 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
912 ir_mode *sp_mode = arch_env->sp->reg_class->mode;
913 dbg_info *dbg = get_irn_dbg_info(free);
914 ir_node *subsp, *mem, *res, *size, *sync;
916 unsigned stack_alignment;
918 /* all non-stack-alloc Free nodes should already be lowered before the
920 assert(get_Free_where(free) == stack_alloc);
922 /* we might need to multiply the size with the element size */
923 if (type != firm_unknown_type && get_type_size_bytes(type) != 1) {
924 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type), mode_Iu);
925 ir_node *cnst = new_rd_Const(dbg, irg, tv);
926 ir_node *mul = new_rd_Mul(dbg, block, get_Free_size(free),
930 size = get_Free_size(free);
933 stack_alignment = 1 << arch_env->stack_alignment;
934 size = adjust_alloc_size(stack_alignment, size, block, dbg);
936 /* The stack pointer will be modified in an unknown manner.
937 We cannot omit it. */
938 env->call->flags.bits.try_omit_fp = 0;
939 subsp = be_new_SubSP(arch_env->sp, block, curr_sp, size);
940 set_irn_dbg_info(subsp, dbg);
942 mem = new_r_Proj(subsp, mode_M, pn_be_SubSP_M);
943 res = new_r_Proj(subsp, sp_mode, pn_be_SubSP_sp);
945 /* we need to sync the memory */
946 in[0] = get_Free_mem(free);
948 sync = new_r_Sync(block, 2, in);
950 /* and make the AddSP dependent on the former memory */
951 add_irn_dep(subsp, get_Free_mem(free));
954 exchange(free, sync);
961 * Check if a node is somehow data dependent on another one.
962 * both nodes must be in the same basic block.
963 * @param n1 The first node.
964 * @param n2 The second node.
965 * @return 1, if n1 is data dependent (transitively) on n2, 0 if not.
967 static int dependent_on(ir_node *n1, ir_node *n2)
969 assert(get_nodes_block(n1) == get_nodes_block(n2));
971 return heights_reachable_in_block(ir_heights, n1, n2);
974 static int cmp_call_dependency(const void *c1, const void *c2)
976 ir_node *n1 = *(ir_node **) c1;
977 ir_node *n2 = *(ir_node **) c2;
981 Classical qsort() comparison function behavior:
982 0 if both elements are equal
983 1 if second is "smaller" that first
984 -1 if first is "smaller" that second
986 if (dependent_on(n1, n2))
989 if (dependent_on(n2, n1))
992 /* The nodes have no depth order, but we need a total order because qsort()
995 * Additionally, we need to respect transitive dependencies. Consider a
996 * Call a depending on Call b and an independent Call c.
997 * We MUST NOT order c > a and b > c. */
998 h1 = get_irn_height(ir_heights, n1);
999 h2 = get_irn_height(ir_heights, n2);
1000 if (h1 < h2) return -1;
1001 if (h1 > h2) return 1;
1002 /* Same height, so use a random (but stable) order */
1003 return get_irn_idx(n1) - get_irn_idx(n2);
1007 * Walker: links all Call/Alloc/Free nodes to the Block they are contained.
1008 * Clears the irg_is_leaf flag if a Call is detected.
1010 static void link_ops_in_block_walker(ir_node *irn, void *data)
1012 be_abi_irg_t *env = (be_abi_irg_t*)data;
1013 unsigned code = get_irn_opcode(irn);
1015 if (code == iro_Call ||
1016 (code == iro_Alloc && get_Alloc_where(irn) == stack_alloc) ||
1017 (code == iro_Free && get_Free_where(irn) == stack_alloc)) {
1018 ir_node *bl = get_nodes_block(irn);
1019 void *save = get_irn_link(bl);
1021 if (code == iro_Call)
1022 env->call->flags.bits.irg_is_leaf = 0;
1024 set_irn_link(irn, save);
1025 set_irn_link(bl, irn);
1028 if (code == iro_Builtin && get_Builtin_kind(irn) == ir_bk_return_address) {
1029 ir_node *param = get_Builtin_param(irn, 0);
1030 ir_tarval *tv = get_Const_tarval(param);
1031 unsigned long value = get_tarval_long(tv);
1032 /* use ebp, so the climbframe algo works... */
1034 env->call->flags.bits.try_omit_fp = 0;
1041 * Process all Call/Alloc/Free nodes inside a basic block.
1042 * Note that the link field of the block must contain a linked list of all
1043 * nodes inside the Block. We first order this list according to data dependency
1044 * and that connect the nodes together.
1046 static void process_ops_in_block(ir_node *bl, void *data)
1048 be_abi_irg_t *env = (be_abi_irg_t*)data;
1049 ir_node *curr_sp = env->init_sp;
1056 for (irn = (ir_node*)get_irn_link(bl); irn != NULL;
1057 irn = (ir_node*)get_irn_link(irn)) {
1061 nodes = ALLOCAN(ir_node*, n_nodes);
1062 for (irn = (ir_node*)get_irn_link(bl), n = 0; irn != NULL;
1063 irn = (ir_node*)get_irn_link(irn), ++n) {
1067 /* If there were call nodes in the block. */
1072 /* order the call nodes according to data dependency */
1073 qsort(nodes, n_nodes, sizeof(nodes[0]), cmp_call_dependency);
1075 for (i = n_nodes - 1; i >= 0; --i) {
1076 ir_node *irn = nodes[i];
1078 DBG((dbg, LEVEL_3, "\tprocessing call %+F\n", irn));
1079 switch (get_irn_opcode(irn)) {
1082 /* The stack pointer will be modified due to a call. */
1083 env->call->flags.bits.try_omit_fp = 0;
1085 curr_sp = adjust_call(env, irn, curr_sp);
1088 if (get_Alloc_where(irn) == stack_alloc)
1089 curr_sp = adjust_alloc(env, irn, curr_sp);
1092 if (get_Free_where(irn) == stack_alloc)
1093 curr_sp = adjust_free(env, irn, curr_sp);
1096 panic("invalid call");
1100 /* Keep the last stack state in the block by tying it to Keep node,
1101 * the proj from calls is already kept */
1102 if (curr_sp != env->init_sp &&
1103 !(is_Proj(curr_sp) && be_is_Call(get_Proj_pred(curr_sp)))) {
1105 keep = be_new_Keep(bl, 1, nodes);
1106 pmap_insert(env->keep_map, bl, keep);
1110 set_irn_link(bl, curr_sp);
1114 * Adjust all call nodes in the graph to the ABI conventions.
1116 static void process_calls(ir_graph *irg)
1118 be_abi_irg_t *abi = be_get_irg_abi(irg);
1120 abi->call->flags.bits.irg_is_leaf = 1;
1121 irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, abi);
1123 ir_heights = heights_new(irg);
1124 irg_block_walk_graph(irg, NULL, process_ops_in_block, abi);
1125 heights_free(ir_heights);
1129 * Computes the stack argument layout type.
1130 * Changes a possibly allocated value param type by moving
1131 * entities to the stack layout type.
1133 * @param env the ABI environment
1134 * @param call the current call ABI
1135 * @param method_type the method type
1136 * @param val_param_tp the value parameter type, will be destroyed
1137 * @param param_map an array mapping method arguments to the stack layout type
1139 * @return the stack argument layout type
1141 static ir_type *compute_arg_type(be_abi_irg_t *env, ir_graph *irg,
1142 be_abi_call_t *call,
1143 ir_type *method_type, ir_type *val_param_tp,
1144 ir_entity ***param_map)
1146 int dir = env->call->flags.bits.left_to_right ? 1 : -1;
1148 int n = get_method_n_params(method_type);
1149 int curr = inc > 0 ? 0 : n - 1;
1150 struct obstack *obst = be_get_be_obst(irg);
1156 ident *id = get_entity_ident(get_irg_entity(irg));
1159 *param_map = map = OALLOCN(obst, ir_entity*, n);
1160 res = new_type_struct(id_mangle_u(id, new_id_from_chars("arg_type", 8)));
1161 for (i = 0; i < n; ++i, curr += inc) {
1162 ir_type *param_type = get_method_param_type(method_type, curr);
1163 be_abi_call_arg_t *arg = get_call_arg(call, 0, curr, 1);
1166 if (arg->on_stack) {
1167 if (val_param_tp != NULL) {
1168 /* the entity was already created, create a copy in the param type */
1169 ir_entity *val_ent = get_method_value_param_ent(method_type, i);
1170 arg->stack_ent = copy_entity_own(val_ent, res);
1171 set_entity_link(val_ent, arg->stack_ent);
1172 set_entity_link(arg->stack_ent, NULL);
1174 /* create a new entity */
1175 snprintf(buf, sizeof(buf), "param_%d", i);
1176 arg->stack_ent = new_entity(res, new_id_from_str(buf), param_type);
1178 ofs += arg->space_before;
1179 ofs = round_up2(ofs, arg->alignment);
1180 set_entity_offset(arg->stack_ent, ofs);
1181 ofs += arg->space_after;
1182 ofs += get_type_size_bytes(param_type);
1183 map[i] = arg->stack_ent;
1186 set_type_size_bytes(res, ofs);
1187 set_type_state(res, layout_fixed);
1192 const arch_register_t *reg;
1196 static int cmp_regs(const void *a, const void *b)
1198 const reg_node_map_t *p = (const reg_node_map_t*)a;
1199 const reg_node_map_t *q = (const reg_node_map_t*)b;
1201 if (p->reg->reg_class == q->reg->reg_class)
1202 return p->reg->index - q->reg->index;
1204 return p->reg->reg_class < q->reg->reg_class ? -1 : +1;
1207 static void reg_map_to_arr(reg_node_map_t *res, pmap *reg_map)
1210 size_t n = pmap_count(reg_map);
1213 foreach_pmap(reg_map, ent) {
1214 res[i].reg = (const arch_register_t*)ent->key;
1215 res[i].irn = (ir_node*)ent->value;
1219 qsort(res, n, sizeof(res[0]), cmp_regs);
1223 * Creates a be_Return for a Return node.
1225 * @param @env the abi environment
1226 * @param irn the Return node or NULL if there was none
1227 * @param bl the block where the be_Retun should be placed
1228 * @param mem the current memory
1229 * @param n_res number of return results
1231 static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl,
1232 ir_node *mem, int n_res)
1234 be_abi_call_t *call = env->call;
1235 ir_graph *irg = get_Block_irg(bl);
1236 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1238 pmap *reg_map = pmap_create();
1239 ir_node *keep = (ir_node*)pmap_get(env->keep_map, bl);
1246 const arch_register_t **regs;
1250 get the valid stack node in this block.
1251 If we had a call in that block there is a Keep constructed by process_calls()
1252 which points to the last stack modification in that block. we'll use
1253 it then. Else we use the stack from the start block and let
1254 the ssa construction fix the usage.
1256 stack = be_abi_reg_map_get(env->regs, arch_env->sp);
1258 stack = get_irn_n(keep, 0);
1260 remove_End_keepalive(get_irg_end(irg), keep);
1263 /* Insert results for Return into the register map. */
1264 for (i = 0; i < n_res; ++i) {
1265 ir_node *res = get_Return_res(irn, i);
1266 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1267 assert(arg->in_reg && "return value must be passed in register");
1268 pmap_insert(reg_map, (void *) arg->reg, res);
1271 /* Add uses of the callee save registers. */
1272 foreach_pmap(env->regs, ent) {
1273 const arch_register_t *reg = (const arch_register_t*)ent->key;
1274 if (reg->type & (arch_register_type_callee_save | arch_register_type_ignore))
1275 pmap_insert(reg_map, ent->key, ent->value);
1278 be_abi_reg_map_set(reg_map, arch_env->sp, stack);
1281 Maximum size of the in array for Return nodes is
1282 return args + callee save/ignore registers + memory + stack pointer
1284 in_max = pmap_count(reg_map) + n_res + 2;
1286 in = ALLOCAN(ir_node*, in_max);
1287 regs = ALLOCAN(arch_register_t const*, in_max);
1290 in[1] = be_abi_reg_map_get(reg_map, arch_env->sp);
1292 regs[1] = arch_env->sp;
1295 /* clear SP entry, since it has already been grown. */
1296 pmap_insert(reg_map, (void *) arch_env->sp, NULL);
1297 for (i = 0; i < n_res; ++i) {
1298 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1300 in[n] = be_abi_reg_map_get(reg_map, arg->reg);
1301 regs[n++] = arg->reg;
1303 /* Clear the map entry to mark the register as processed. */
1304 be_abi_reg_map_set(reg_map, arg->reg, NULL);
1307 /* grow the rest of the stuff. */
1308 foreach_pmap(reg_map, ent) {
1310 in[n] = (ir_node*)ent->value;
1311 regs[n++] = (const arch_register_t*)ent->key;
1315 /* The in array for the new back end return is now ready. */
1317 dbgi = get_irn_dbg_info(irn);
1321 /* we have to pop the shadow parameter in in case of struct returns */
1323 ret = be_new_Return(dbgi, irg, bl, n_res, pop, n, in);
1325 /* Set the register classes of the return's parameter accordingly. */
1326 for (i = 0; i < n; ++i) {
1327 if (regs[i] == NULL)
1330 be_set_constr_single_reg_in(ret, i, regs[i], arch_register_req_type_none);
1333 /* Free the space of the Epilog's in array and the register <-> proj map. */
1334 pmap_destroy(reg_map);
1339 typedef struct ent_pos_pair ent_pos_pair;
1340 struct ent_pos_pair {
1341 ir_entity *ent; /**< a value param entity */
1342 int pos; /**< its parameter number */
1343 ent_pos_pair *next; /**< for linking */
1346 typedef struct lower_frame_sels_env_t {
1347 ent_pos_pair *value_param_list; /**< the list of all value param entities */
1348 ir_node *frame; /**< the current frame */
1349 const arch_register_class_t *sp_class; /**< register class of the stack pointer */
1350 const arch_register_class_t *link_class; /**< register class of the link pointer */
1351 ir_type *value_tp; /**< the value type if any */
1352 ir_type *frame_tp; /**< the frame type */
1353 int static_link_pos; /**< argument number of the hidden static link */
1354 } lower_frame_sels_env_t;
1357 * Return an entity from the backend for an value param entity.
1359 * @param ent an value param type entity
1360 * @param ctx context
1362 static ir_entity *get_argument_entity(ir_entity *ent, lower_frame_sels_env_t *ctx)
1364 ir_entity *argument_ent = (ir_entity*)get_entity_link(ent);
1366 if (argument_ent == NULL) {
1367 /* we have NO argument entity yet: This is bad, as we will
1368 * need one for backing store.
1371 ir_type *frame_tp = ctx->frame_tp;
1372 unsigned offset = get_type_size_bytes(frame_tp);
1373 ir_type *tp = get_entity_type(ent);
1374 unsigned align = get_type_alignment_bytes(tp);
1376 offset += align - 1;
1377 offset &= ~(align - 1);
1379 argument_ent = copy_entity_own(ent, frame_tp);
1381 /* must be automatic to set a fixed layout */
1382 set_entity_offset(argument_ent, offset);
1383 offset += get_type_size_bytes(tp);
1385 set_type_size_bytes(frame_tp, offset);
1386 set_entity_link(ent, argument_ent);
1388 return argument_ent;
1391 * Walker: Replaces Sels of frame type and
1392 * value param type entities by FrameAddress.
1393 * Links all used entities.
1395 static void lower_frame_sels_walker(ir_node *irn, void *data)
1397 lower_frame_sels_env_t *ctx = (lower_frame_sels_env_t*)data;
1400 ir_node *ptr = get_Sel_ptr(irn);
1402 if (ptr == ctx->frame) {
1403 ir_entity *ent = get_Sel_entity(irn);
1404 ir_node *bl = get_nodes_block(irn);
1407 int is_value_param = 0;
1409 if (get_entity_owner(ent) == ctx->value_tp) {
1412 /* replace by its copy from the argument type */
1413 pos = get_struct_member_index(ctx->value_tp, ent);
1414 ent = get_argument_entity(ent, ctx);
1417 nw = be_new_FrameAddr(ctx->sp_class, bl, ctx->frame, ent);
1420 /* check, if it's a param Sel and if have not seen this entity before */
1421 if (is_value_param && get_entity_link(ent) == NULL) {
1427 ARR_APP1(ent_pos_pair, ctx->value_param_list, pair);
1429 set_entity_link(ent, ctx->value_param_list);
1436 * Check if a value parameter is transmitted as a register.
1437 * This might happen if the address of an parameter is taken which is
1438 * transmitted in registers.
1440 * Note that on some architectures this case must be handled specially
1441 * because the place of the backing store is determined by their ABI.
1443 * In the default case we move the entity to the frame type and create
1444 * a backing store into the first block.
1446 static void fix_address_of_parameter_access(be_abi_irg_t *env, ir_graph *irg,
1447 ent_pos_pair *value_param_list)
1449 be_abi_call_t *call = env->call;
1450 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1451 ent_pos_pair *entry, *new_list;
1453 int i, n = ARR_LEN(value_param_list);
1456 for (i = 0; i < n; ++i) {
1457 int pos = value_param_list[i].pos;
1458 be_abi_call_arg_t *arg = get_call_arg(call, 0, pos, 1);
1461 DBG((dbg, LEVEL_2, "\targ #%d need backing store\n", pos));
1462 value_param_list[i].next = new_list;
1463 new_list = &value_param_list[i];
1466 if (new_list != NULL) {
1467 /* ok, change the graph */
1468 ir_node *start_bl = get_irg_start_block(irg);
1469 ir_node *first_bl = get_first_block_succ(start_bl);
1470 ir_node *frame, *imem, *nmem, *store, *mem, *args;
1471 optimization_state_t state;
1474 assert(first_bl && first_bl != start_bl);
1475 /* we had already removed critical edges, so the following
1476 assertion should be always true. */
1477 assert(get_Block_n_cfgpreds(first_bl) == 1);
1479 /* now create backing stores */
1480 frame = get_irg_frame(irg);
1481 imem = get_irg_initial_mem(irg);
1483 save_optimization_state(&state);
1485 nmem = new_r_Proj(get_irg_start(irg), mode_M, pn_Start_M);
1486 restore_optimization_state(&state);
1488 /* reroute all edges to the new memory source */
1489 edges_reroute(imem, nmem);
1493 args = get_irg_args(irg);
1494 for (entry = new_list; entry != NULL; entry = entry->next) {
1496 ir_type *tp = get_entity_type(entry->ent);
1497 ir_mode *mode = get_type_mode(tp);
1500 /* address for the backing store */
1501 addr = be_new_FrameAddr(arch_env->sp->reg_class, first_bl, frame, entry->ent);
1504 mem = new_r_Proj(store, mode_M, pn_Store_M);
1506 /* the backing store itself */
1507 store = new_r_Store(first_bl, mem, addr,
1508 new_r_Proj(args, mode, i), cons_none);
1510 /* the new memory Proj gets the last Proj from store */
1511 set_Proj_pred(nmem, store);
1512 set_Proj_proj(nmem, pn_Store_M);
1514 /* move all entities to the frame type */
1515 frame_tp = get_irg_frame_type(irg);
1516 offset = get_type_size_bytes(frame_tp);
1518 /* we will add new entities: set the layout to undefined */
1519 assert(get_type_state(frame_tp) == layout_fixed);
1520 set_type_state(frame_tp, layout_undefined);
1521 for (entry = new_list; entry != NULL; entry = entry->next) {
1522 ir_entity *ent = entry->ent;
1524 /* If the entity is still on the argument type, move it to the
1526 * This happens if the value_param type was build due to compound
1528 if (get_entity_owner(ent) != frame_tp) {
1529 ir_type *tp = get_entity_type(ent);
1530 unsigned align = get_type_alignment_bytes(tp);
1532 offset += align - 1;
1533 offset &= ~(align - 1);
1534 set_entity_owner(ent, frame_tp);
1535 /* must be automatic to set a fixed layout */
1536 set_entity_offset(ent, offset);
1537 offset += get_type_size_bytes(tp);
1540 set_type_size_bytes(frame_tp, offset);
1541 /* fix the layout again */
1542 set_type_state(frame_tp, layout_fixed);
1547 * The start block has no jump, instead it has an initial exec Proj.
1548 * The backend wants to handle all blocks the same way, so we replace
1549 * the out cfg edge with a real jump.
1551 static void fix_start_block(ir_graph *irg)
1553 ir_node *initial_X = get_irg_initial_exec(irg);
1554 ir_node *start_block = get_irg_start_block(irg);
1555 ir_node *jmp = new_r_Jmp(start_block);
1557 assert(is_Proj(initial_X));
1558 exchange(initial_X, jmp);
1559 set_irg_initial_exec(irg, new_r_Bad(irg, mode_X));
1563 * Update the entity of Sels to the outer value parameters.
1565 static void update_outer_frame_sels(ir_node *irn, void *env)
1567 lower_frame_sels_env_t *ctx = (lower_frame_sels_env_t*)env;
1574 ptr = get_Sel_ptr(irn);
1575 if (! is_arg_Proj(ptr))
1577 if (get_Proj_proj(ptr) != ctx->static_link_pos)
1579 ent = get_Sel_entity(irn);
1581 if (get_entity_owner(ent) == ctx->value_tp) {
1582 /* replace by its copy from the argument type */
1583 pos = get_struct_member_index(ctx->value_tp, ent);
1584 ent = get_argument_entity(ent, ctx);
1585 set_Sel_entity(irn, ent);
1587 /* check, if we have not seen this entity before */
1588 if (get_entity_link(ent) == NULL) {
1594 ARR_APP1(ent_pos_pair, ctx->value_param_list, pair);
1596 set_entity_link(ent, ctx->value_param_list);
1602 * Fix access to outer local variables.
1604 static void fix_outer_variable_access(be_abi_irg_t *env,
1605 lower_frame_sels_env_t *ctx)
1611 for (i = get_class_n_members(ctx->frame_tp) - 1; i >= 0; --i) {
1612 ir_entity *ent = get_class_member(ctx->frame_tp, i);
1614 if (! is_method_entity(ent))
1617 irg = get_entity_irg(ent);
1622 * FIXME: find the number of the static link parameter
1623 * for now we assume 0 here
1625 ctx->static_link_pos = 0;
1627 irg_walk_graph(irg, NULL, update_outer_frame_sels, ctx);
1632 * Modify the irg itself and the frame type.
1634 static void modify_irg(ir_graph *irg)
1636 be_abi_irg_t *env = be_get_irg_abi(irg);
1637 be_abi_call_t *call = env->call;
1638 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1639 const arch_register_t *sp = arch_env->sp;
1640 ir_type *method_type = get_entity_type(get_irg_entity(irg));
1641 be_irg_t *birg = be_birg_from_irg(irg);
1642 struct obstack *obst = be_get_be_obst(irg);
1643 be_stack_layout_t *stack_layout = be_get_irg_stack_layout(irg);
1646 ir_node *new_mem_proj;
1652 unsigned frame_size;
1655 const arch_register_t *fp_reg;
1656 ir_node *frame_pointer;
1660 const ir_edge_t *edge;
1661 ir_type *arg_type, *bet_type, *tp;
1662 lower_frame_sels_env_t ctx;
1663 ir_entity **param_map;
1665 DBG((dbg, LEVEL_1, "introducing abi on %+F\n", irg));
1667 old_mem = get_irg_initial_mem(irg);
1669 irp_reserve_resources(irp, IR_RESOURCE_ENTITY_LINK);
1671 /* set the links of all frame entities to NULL, we use it
1672 to detect if an entity is already linked in the value_param_list */
1673 tp = get_method_value_param_type(method_type);
1676 /* clear the links of the clone type, let the
1677 original entities point to its clones */
1678 for (i = get_struct_n_members(tp) - 1; i >= 0; --i) {
1679 ir_entity *mem = get_struct_member(tp, i);
1680 set_entity_link(mem, NULL);
1684 arg_type = compute_arg_type(env, irg, call, method_type, tp, ¶m_map);
1686 /* Convert the Sel nodes in the irg to frame addr nodes: */
1687 ctx.value_param_list = NEW_ARR_F(ent_pos_pair, 0);
1688 ctx.frame = get_irg_frame(irg);
1689 ctx.sp_class = arch_env->sp->reg_class;
1690 ctx.link_class = arch_env->link_class;
1691 ctx.frame_tp = get_irg_frame_type(irg);
1693 /* layout the stackframe now */
1694 if (get_type_state(ctx.frame_tp) == layout_undefined) {
1695 default_layout_compound_type(ctx.frame_tp);
1698 /* we will possible add new entities to the frame: set the layout to undefined */
1699 assert(get_type_state(ctx.frame_tp) == layout_fixed);
1700 set_type_state(ctx.frame_tp, layout_undefined);
1702 irg_walk_graph(irg, lower_frame_sels_walker, NULL, &ctx);
1704 /* fix the frame type layout again */
1705 set_type_state(ctx.frame_tp, layout_fixed);
1706 /* align stackframe to 4 byte */
1707 frame_size = get_type_size_bytes(ctx.frame_tp);
1708 if (frame_size % 4 != 0) {
1709 set_type_size_bytes(ctx.frame_tp, frame_size + 4 - (frame_size % 4));
1712 env->regs = pmap_create();
1714 n_params = get_method_n_params(method_type);
1715 args = OALLOCNZ(obst, ir_node*, n_params);
1718 * for inner function we must now fix access to outer frame entities.
1720 fix_outer_variable_access(env, &ctx);
1722 /* Check if a value parameter is transmitted as a register.
1723 * This might happen if the address of an parameter is taken which is
1724 * transmitted in registers.
1726 * Note that on some architectures this case must be handled specially
1727 * because the place of the backing store is determined by their ABI.
1729 * In the default case we move the entity to the frame type and create
1730 * a backing store into the first block.
1732 fix_address_of_parameter_access(env, irg, ctx.value_param_list);
1734 DEL_ARR_F(ctx.value_param_list);
1735 irp_free_resources(irp, IR_RESOURCE_ENTITY_LINK);
1737 /* Fill the argument vector */
1738 arg_tuple = get_irg_args(irg);
1739 foreach_out_edge(arg_tuple, edge) {
1740 ir_node *irn = get_edge_src_irn(edge);
1741 if (! is_Anchor(irn)) {
1742 int nr = get_Proj_proj(irn);
1744 DBG((dbg, LEVEL_2, "\treading arg: %d -> %+F\n", nr, irn));
1748 stack_layout->sp_relative = call->flags.bits.try_omit_fp;
1749 bet_type = call->cb->get_between_type(irg);
1750 stack_frame_init(stack_layout, arg_type, bet_type,
1751 get_irg_frame_type(irg), param_map);
1753 /* Count the register params and add them to the number of Projs for the RegParams node */
1754 for (i = 0; i < n_params; ++i) {
1755 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1756 if (arg->in_reg && args[i]) {
1757 assert(arg->reg != sp && "cannot use stack pointer as parameter register");
1758 assert(i == get_Proj_proj(args[i]));
1760 /* For now, associate the register with the old Proj from Start representing that argument. */
1761 pmap_insert(env->regs, (void *) arg->reg, args[i]);
1762 DBG((dbg, LEVEL_2, "\targ #%d -> reg %s\n", i, arg->reg->name));
1766 /* Collect all callee-save registers */
1767 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
1768 const arch_register_class_t *cls = &arch_env->register_classes[i];
1769 for (j = 0; j < cls->n_regs; ++j) {
1770 const arch_register_t *reg = &cls->regs[j];
1771 if (reg->type & (arch_register_type_callee_save | arch_register_type_state)) {
1772 pmap_insert(env->regs, (void *) reg, NULL);
1777 fp_reg = call->flags.bits.try_omit_fp ? arch_env->sp : arch_env->bp;
1778 rbitset_clear(birg->allocatable_regs, fp_reg->global_index);
1780 /* handle start block here (place a jump in the block) */
1781 fix_start_block(irg);
1783 pmap_insert(env->regs, (void *) sp, NULL);
1784 pmap_insert(env->regs, (void *) arch_env->bp, NULL);
1785 start_bl = get_irg_start_block(irg);
1786 env->start = be_new_Start(NULL, start_bl, pmap_count(env->regs) + 1);
1787 set_irg_start(irg, env->start);
1790 * make proj nodes for the callee save registers.
1791 * memorize them, since Return nodes get those as inputs.
1793 * Note, that if a register corresponds to an argument, the regs map
1794 * contains the old Proj from start for that argument.
1796 rm = ALLOCAN(reg_node_map_t, pmap_count(env->regs));
1797 reg_map_to_arr(rm, env->regs);
1798 for (i = 0, n = pmap_count(env->regs); i < n; ++i) {
1799 const arch_register_t *reg = rm[i].reg;
1800 ir_mode *mode = reg->reg_class->mode;
1802 arch_register_req_type_t add_type = arch_register_req_type_none;
1806 add_type |= arch_register_req_type_produces_sp;
1807 if (!rbitset_is_set(birg->allocatable_regs, reg->global_index)) {
1808 add_type |= arch_register_req_type_ignore;
1812 proj = new_r_Proj(env->start, mode, nr + 1);
1813 pmap_insert(env->regs, (void *) reg, proj);
1814 be_set_constr_single_reg_out(env->start, nr + 1, reg, add_type);
1815 arch_set_irn_register(proj, reg);
1817 DBG((dbg, LEVEL_2, "\tregister save proj #%d -> reg %s\n", nr, reg->name));
1820 /* create a new initial memory proj */
1821 assert(is_Proj(old_mem));
1822 arch_set_out_register_req(env->start, 0, arch_no_register_req);
1823 new_mem_proj = new_r_Proj(env->start, mode_M, 0);
1825 set_irg_initial_mem(irg, mem);
1827 env->init_sp = be_abi_reg_map_get(env->regs, sp);
1829 /* set new frame_pointer */
1830 frame_pointer = be_abi_reg_map_get(env->regs, fp_reg);
1831 set_irg_frame(irg, frame_pointer);
1833 /* rewire old mem users to new mem */
1834 exchange(old_mem, mem);
1836 /* keep the mem (for functions with an endless loop = no return) */
1839 set_irg_initial_mem(irg, mem);
1841 /* Now, introduce stack param nodes for all parameters passed on the stack */
1842 for (i = 0; i < n_params; ++i) {
1843 ir_node *arg_proj = args[i];
1844 ir_node *repl = NULL;
1846 if (arg_proj != NULL) {
1847 be_abi_call_arg_t *arg;
1848 ir_type *param_type;
1849 int nr = get_Proj_proj(arg_proj);
1852 nr = MIN(nr, n_params);
1853 arg = get_call_arg(call, 0, nr, 1);
1854 param_type = get_method_param_type(method_type, nr);
1857 repl = (ir_node*)pmap_get(env->regs, arg->reg);
1858 } else if (arg->on_stack) {
1859 ir_node *addr = be_new_FrameAddr(sp->reg_class, start_bl, frame_pointer, arg->stack_ent);
1861 /* For atomic parameters which are actually used, we create a Load node. */
1862 if (is_atomic_type(param_type) && get_irn_n_edges(args[i]) > 0) {
1863 ir_mode *mode = get_type_mode(param_type);
1864 ir_mode *load_mode = arg->load_mode;
1865 ir_node *nomem = get_irg_no_mem(irg);
1867 ir_node *load = new_r_Load(start_bl, nomem, addr, load_mode, cons_floats);
1868 repl = new_r_Proj(load, load_mode, pn_Load_res);
1870 if (mode != load_mode) {
1871 repl = new_r_Conv(start_bl, repl, mode);
1874 /* The stack parameter is not primitive (it is a struct or array),
1875 * we thus will create a node representing the parameter's address
1881 assert(repl != NULL);
1883 /* Beware: the mode of the register parameters is always the mode of the register class
1884 which may be wrong. Add Conv's then. */
1885 mode = get_irn_mode(args[i]);
1886 if (mode != get_irn_mode(repl)) {
1887 repl = new_r_Conv(get_nodes_block(repl), repl, mode);
1889 exchange(args[i], repl);
1893 /* the arg proj is not needed anymore now and should be only used by the anchor */
1894 assert(get_irn_n_edges(arg_tuple) == 1);
1895 kill_node(arg_tuple);
1896 set_irg_args(irg, new_r_Bad(irg, mode_T));
1898 /* All Return nodes hang on the End node, so look for them there. */
1899 end = get_irg_end_block(irg);
1900 for (i = 0, n = get_Block_n_cfgpreds(end); i < n; ++i) {
1901 ir_node *irn = get_Block_cfgpred(end, i);
1903 if (is_Return(irn)) {
1904 ir_node *blk = get_nodes_block(irn);
1905 ir_node *mem = get_Return_mem(irn);
1906 ir_node *ret = create_be_return(env, irn, blk, mem, get_Return_n_ress(irn));
1911 /* if we have endless loops here, n might be <= 0. Do NOT create a be_Return then,
1912 the code is dead and will never be executed. */
1915 /** Fix the state inputs of calls that still hang on unknowns */
1916 static void fix_call_state_inputs(ir_graph *irg)
1918 be_abi_irg_t *env = be_get_irg_abi(irg);
1919 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1921 arch_register_t **stateregs = NEW_ARR_F(arch_register_t*, 0);
1923 /* Collect caller save registers */
1924 n = arch_env->n_register_classes;
1925 for (i = 0; i < n; ++i) {
1927 const arch_register_class_t *cls = &arch_env->register_classes[i];
1928 for (j = 0; j < cls->n_regs; ++j) {
1929 const arch_register_t *reg = arch_register_for_index(cls, j);
1930 if (reg->type & arch_register_type_state) {
1931 ARR_APP1(arch_register_t*, stateregs, (arch_register_t *)reg);
1936 n = ARR_LEN(env->calls);
1937 n_states = ARR_LEN(stateregs);
1938 for (i = 0; i < n; ++i) {
1940 ir_node *call = env->calls[i];
1942 arity = get_irn_arity(call);
1944 /* the state reg inputs are the last n inputs of the calls */
1945 for (s = 0; s < n_states; ++s) {
1946 int inp = arity - n_states + s;
1947 const arch_register_t *reg = stateregs[s];
1948 ir_node *regnode = be_abi_reg_map_get(env->regs, reg);
1950 set_irn_n(call, inp, regnode);
1954 DEL_ARR_F(stateregs);
1958 * Create a trampoline entity for the given method.
1960 static ir_entity *create_trampoline(be_main_env_t *be, ir_entity *method)
1962 ir_type *type = get_entity_type(method);
1963 ident *old_id = get_entity_ld_ident(method);
1964 ident *id = id_mangle3("", old_id, "$stub");
1965 ir_type *parent = be->pic_trampolines_type;
1966 ir_entity *ent = new_entity(parent, old_id, type);
1967 set_entity_ld_ident(ent, id);
1968 set_entity_visibility(ent, ir_visibility_private);
1974 * Returns the trampoline entity for the given method.
1976 static ir_entity *get_trampoline(be_main_env_t *env, ir_entity *method)
1978 ir_entity *result = (ir_entity*)pmap_get(env->ent_trampoline_map, method);
1979 if (result == NULL) {
1980 result = create_trampoline(env, method);
1981 pmap_insert(env->ent_trampoline_map, method, result);
1987 static ir_entity *create_pic_symbol(be_main_env_t *be, ir_entity *entity)
1989 ident *old_id = get_entity_ld_ident(entity);
1990 ident *id = id_mangle3("", old_id, "$non_lazy_ptr");
1991 ir_type *e_type = get_entity_type(entity);
1992 ir_type *type = new_type_pointer(e_type);
1993 ir_type *parent = be->pic_symbols_type;
1994 ir_entity *ent = new_entity(parent, old_id, type);
1995 set_entity_ld_ident(ent, id);
1996 set_entity_visibility(ent, ir_visibility_private);
2001 static ir_entity *get_pic_symbol(be_main_env_t *env, ir_entity *entity)
2003 ir_entity *result = (ir_entity*)pmap_get(env->ent_pic_symbol_map, entity);
2004 if (result == NULL) {
2005 result = create_pic_symbol(env, entity);
2006 pmap_insert(env->ent_pic_symbol_map, entity, result);
2015 * Returns non-zero if a given entity can be accessed using a relative address.
2017 static int can_address_relative(ir_entity *entity)
2019 return get_entity_visibility(entity) != ir_visibility_external
2020 && !(get_entity_linkage(entity) & IR_LINKAGE_MERGE);
2023 static ir_node *get_pic_base(ir_graph *irg)
2025 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
2026 if (arch_env->impl->get_pic_base == NULL)
2028 return arch_env->impl->get_pic_base(irg);
2031 /** patches SymConsts to work in position independent code */
2032 static void fix_pic_symconsts(ir_node *node, void *data)
2034 ir_graph *irg = get_irn_irg(node);
2035 be_main_env_t *be = be_get_irg_main_env(irg);
2045 arity = get_irn_arity(node);
2046 for (i = 0; i < arity; ++i) {
2048 ir_node *pred = get_irn_n(node, i);
2050 ir_entity *pic_symbol;
2051 ir_node *pic_symconst;
2053 if (!is_SymConst(pred))
2056 entity = get_SymConst_entity(pred);
2057 block = get_nodes_block(pred);
2059 /* calls can jump to relative addresses, so we can directly jump to
2060 the (relatively) known call address or the trampoline */
2061 if (i == 1 && is_Call(node)) {
2062 ir_entity *trampoline;
2063 ir_node *trampoline_const;
2065 if (can_address_relative(entity))
2068 dbgi = get_irn_dbg_info(pred);
2069 trampoline = get_trampoline(be, entity);
2070 trampoline_const = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
2072 set_irn_n(node, i, trampoline_const);
2076 /* everything else is accessed relative to EIP */
2077 mode = get_irn_mode(pred);
2078 pic_base = get_pic_base(irg);
2080 /* all ok now for locally constructed stuff */
2081 if (can_address_relative(entity)) {
2082 ir_node *add = new_r_Add(block, pic_base, pred, mode);
2084 /* make sure the walker doesn't visit this add again */
2085 mark_irn_visited(add);
2086 set_irn_n(node, i, add);
2090 /* get entry from pic symbol segment */
2091 dbgi = get_irn_dbg_info(pred);
2092 pic_symbol = get_pic_symbol(be, entity);
2093 pic_symconst = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
2095 add = new_r_Add(block, pic_base, pic_symconst, mode);
2096 mark_irn_visited(add);
2098 /* we need an extra indirection for global data outside our current
2099 module. The loads are always safe and can therefore float
2100 and need no memory input */
2101 load = new_r_Load(block, get_irg_no_mem(irg), add, mode, cons_floats);
2102 load_res = new_r_Proj(load, mode, pn_Load_res);
2104 set_irn_n(node, i, load_res);
2108 void be_abi_introduce(ir_graph *irg)
2110 be_abi_irg_t *env = XMALLOCZ(be_abi_irg_t);
2111 ir_node *old_frame = get_irg_frame(irg);
2112 be_options_t *options = be_get_irg_options(irg);
2113 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
2114 ir_entity *entity = get_irg_entity(irg);
2115 ir_type *method_type = get_entity_type(entity);
2116 be_irg_t *birg = be_birg_from_irg(irg);
2117 struct obstack *obst = &birg->obst;
2118 ir_node *dummy = new_r_Dummy(irg,
2119 arch_env->sp->reg_class->mode);
2122 /* determine allocatable registers */
2123 assert(birg->allocatable_regs == NULL);
2124 birg->allocatable_regs = rbitset_obstack_alloc(obst, arch_env->n_registers);
2125 for (r = 0; r < arch_env->n_registers; ++r) {
2126 const arch_register_t *reg = &arch_env->registers[r];
2127 if ( !(reg->type & arch_register_type_ignore)) {
2128 rbitset_set(birg->allocatable_regs, r);
2132 /* break here if backend provides a custom API.
2133 * Note: we shouldn't have to setup any be_abi_irg_t* stuff at all,
2134 * but need more cleanup to make this work
2136 be_set_irg_abi(irg, env);
2138 be_omit_fp = options->omit_fp;
2140 env->keep_map = pmap_create();
2141 env->call = be_abi_call_new(arch_env->sp->reg_class);
2142 arch_env_get_call_abi(arch_env, method_type, env->call);
2144 env->init_sp = dummy;
2145 env->calls = NEW_ARR_F(ir_node*, 0);
2150 irg_walk_graph(irg, fix_pic_symconsts, NULL, env);
2153 /* Lower all call nodes in the IRG. */
2156 /* Process the IRG */
2159 /* fix call inputs for state registers */
2160 fix_call_state_inputs(irg);
2162 /* We don't need the keep map anymore. */
2163 pmap_destroy(env->keep_map);
2164 env->keep_map = NULL;
2166 /* calls array is not needed anymore */
2167 DEL_ARR_F(env->calls);
2170 /* reroute the stack origin of the calls to the true stack origin. */
2171 exchange(dummy, env->init_sp);
2172 exchange(old_frame, get_irg_frame(irg));
2174 pmap_destroy(env->regs);
2178 void be_abi_free(ir_graph *irg)
2180 be_abi_irg_t *env = be_get_irg_abi(irg);
2182 if (env->call != NULL)
2183 be_abi_call_free(env->call);
2184 assert(env->regs == NULL);
2187 be_set_irg_abi(irg, NULL);
2190 void be_put_allocatable_regs(const ir_graph *irg,
2191 const arch_register_class_t *cls, bitset_t *bs)
2193 be_irg_t *birg = be_birg_from_irg(irg);
2194 unsigned *allocatable_regs = birg->allocatable_regs;
2197 assert(bitset_size(bs) == cls->n_regs);
2198 bitset_clear_all(bs);
2199 for (i = 0; i < cls->n_regs; ++i) {
2200 const arch_register_t *reg = &cls->regs[i];
2201 if (rbitset_is_set(allocatable_regs, reg->global_index))
2206 unsigned be_get_n_allocatable_regs(const ir_graph *irg,
2207 const arch_register_class_t *cls)
2209 bitset_t *bs = bitset_alloca(cls->n_regs);
2210 be_put_allocatable_regs(irg, cls, bs);
2211 return bitset_popcount(bs);
2214 void be_set_allocatable_regs(const ir_graph *irg,
2215 const arch_register_class_t *cls,
2216 unsigned *raw_bitset)
2218 be_irg_t *birg = be_birg_from_irg(irg);
2219 unsigned *allocatable_regs = birg->allocatable_regs;
2222 rbitset_clear_all(raw_bitset, cls->n_regs);
2223 for (i = 0; i < cls->n_regs; ++i) {
2224 const arch_register_t *reg = &cls->regs[i];
2225 if (rbitset_is_set(allocatable_regs, reg->global_index))
2226 rbitset_set(raw_bitset, i);
2230 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_abi)
2231 void be_init_abi(void)
2233 FIRM_DBG_REGISTER(dbg, "firm.be.abi");