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 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"
59 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
61 typedef struct be_abi_call_arg_t {
62 unsigned is_res : 1; /**< 1: the call argument is a return value. 0: it's a call parameter. */
63 unsigned in_reg : 1; /**< 1: this argument is transmitted in registers. */
64 unsigned on_stack : 1; /**< 1: this argument is transmitted on the stack. */
65 unsigned callee : 1; /**< 1: someone called us. 0: We call another function */
68 const arch_register_t *reg;
71 unsigned alignment; /**< stack alignment */
72 unsigned space_before; /**< allocate space before */
73 unsigned space_after; /**< allocate space after */
76 struct be_abi_call_t {
77 be_abi_call_flags_t flags; /**< Flags describing the ABI behavior on calls */
78 int pop; /**< number of bytes the stack frame is shrinked by the callee on return. */
79 const be_abi_callbacks_t *cb;
80 ir_type *between_type;
82 const arch_register_class_t *cls_addr; /**< register class of the call address */
86 * The ABI information for the current graph.
89 survive_dce_t *dce_survivor;
91 be_abi_call_t *call; /**< The ABI call information. */
93 ir_node *init_sp; /**< The node representing the stack pointer
94 at the start of the function. */
96 ir_node *start; /**< The be_Start params node. */
97 pmap *regs; /**< A map of all callee-save and ignore regs to
98 their Projs to the RegParams node. */
100 int start_block_bias; /**< The stack bias at the end of the start block. */
102 void *cb; /**< ABI Callback self pointer. */
104 pmap *keep_map; /**< mapping blocks to keep nodes. */
106 ir_node **calls; /**< flexible array containing all be_Call nodes */
109 static ir_heights_t *ir_heights;
111 /** Flag: if set, try to omit the frame pointer in all routines. */
112 static int be_omit_fp = 1;
115 _ ____ ___ ____ _ _ _ _
116 / \ | __ )_ _| / ___|__ _| | | |__ __ _ ___| | _____
117 / _ \ | _ \| | | | / _` | | | '_ \ / _` |/ __| |/ / __|
118 / ___ \| |_) | | | |__| (_| | | | |_) | (_| | (__| <\__ \
119 /_/ \_\____/___| \____\__,_|_|_|_.__/ \__,_|\___|_|\_\___/
121 These callbacks are used by the backend to set the parameters
122 for a specific call type.
126 * Set compare function: compares two ABI call object arguments.
128 static int cmp_call_arg(const void *a, const void *b, size_t n)
130 const be_abi_call_arg_t *p = a, *q = b;
132 return !(p->is_res == q->is_res && p->pos == q->pos && p->callee == q->callee);
136 * Get an ABI call object argument.
138 * @param call the abi call
139 * @param is_res true for call results, false for call arguments
140 * @param pos position of the argument
141 * @param callee context type - if we are callee or caller
143 static be_abi_call_arg_t *get_call_arg(be_abi_call_t *call, int is_res, int pos, int callee)
145 be_abi_call_arg_t arg;
148 memset(&arg, 0, sizeof(arg));
153 hash = is_res * 128 + pos;
155 return set_find(call->params, &arg, sizeof(arg), hash);
159 * Set an ABI call object argument.
161 static void remember_call_arg(be_abi_call_arg_t *arg, be_abi_call_t *call, be_abi_context_t context)
163 unsigned hash = arg->is_res * 128 + arg->pos;
164 if (context & ABI_CONTEXT_CALLEE) {
166 set_insert(call->params, arg, sizeof(*arg), hash);
168 if (context & ABI_CONTEXT_CALLER) {
170 set_insert(call->params, arg, sizeof(*arg), hash);
174 /* Set the flags for a call. */
175 void be_abi_call_set_flags(be_abi_call_t *call, be_abi_call_flags_t flags, const be_abi_callbacks_t *cb)
181 /* Sets the number of bytes the stackframe is shrinked by the callee on return */
182 void be_abi_call_set_pop(be_abi_call_t *call, int pop)
188 /* Set register class for call address */
189 void be_abi_call_set_call_address_reg_class(be_abi_call_t *call, const arch_register_class_t *cls)
191 call->cls_addr = cls;
195 void be_abi_call_param_stack(be_abi_call_t *call, int arg_pos,
196 ir_mode *load_mode, unsigned alignment,
197 unsigned space_before, unsigned space_after,
198 be_abi_context_t context)
200 be_abi_call_arg_t arg;
201 memset(&arg, 0, sizeof(arg));
202 assert(alignment > 0 && "Alignment must be greater than 0");
204 arg.load_mode = load_mode;
205 arg.alignment = alignment;
206 arg.space_before = space_before;
207 arg.space_after = space_after;
211 remember_call_arg(&arg, call, context);
214 void be_abi_call_param_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
216 be_abi_call_arg_t arg;
217 memset(&arg, 0, sizeof(arg));
224 remember_call_arg(&arg, call, context);
227 void be_abi_call_res_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
229 be_abi_call_arg_t arg;
230 memset(&arg, 0, sizeof(arg));
237 remember_call_arg(&arg, call, context);
240 /* Get the flags of a ABI call object. */
241 be_abi_call_flags_t be_abi_call_get_flags(const be_abi_call_t *call)
247 * Constructor for a new ABI call object.
249 * @param cls_addr register class of the call address
251 * @return the new ABI call object
253 static be_abi_call_t *be_abi_call_new(const arch_register_class_t *cls_addr)
255 be_abi_call_t *call = XMALLOCZ(be_abi_call_t);
258 call->params = new_set(cmp_call_arg, 16);
260 call->cls_addr = cls_addr;
262 call->flags.bits.try_omit_fp = be_omit_fp;
268 * Destructor for an ABI call object.
270 static void be_abi_call_free(be_abi_call_t *call)
272 del_set(call->params);
277 * Initializes the frame layout from parts
279 * @param frame the stack layout that will be initialized
280 * @param args the stack argument layout type
281 * @param between the between layout type
282 * @param locals the method frame type
283 * @param stack_dir the stack direction: < 0 decreasing, > 0 increasing addresses
284 * @param param_map an array mapping method argument positions to the stack argument type
286 * @return the initialized stack layout
288 static be_stack_layout_t *stack_frame_init(be_stack_layout_t *frame, ir_type *args,
289 ir_type *between, ir_type *locals, int stack_dir,
290 ir_entity *param_map[])
292 frame->arg_type = args;
293 frame->between_type = between;
294 frame->frame_type = locals;
295 frame->initial_offset = 0;
296 frame->initial_bias = 0;
297 frame->stack_dir = stack_dir;
298 frame->order[1] = between;
299 frame->param_map = param_map;
302 frame->order[0] = args;
303 frame->order[2] = locals;
305 /* typical decreasing stack: locals have the
306 * lowest addresses, arguments the highest */
307 frame->order[0] = locals;
308 frame->order[2] = args;
320 Adjustment of the calls inside a graph.
325 * Transform a call node into a be_Call node.
327 * @param env The ABI environment for the current irg.
328 * @param irn The call node.
329 * @param curr_sp The stack pointer node to use.
330 * @return The stack pointer after the call.
332 static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
334 ir_graph *irg = get_irn_irg(irn);
335 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
336 ir_type *call_tp = get_Call_type(irn);
337 ir_node *call_ptr = get_Call_ptr(irn);
338 int n_params = get_method_n_params(call_tp);
339 ir_node *curr_mem = get_Call_mem(irn);
340 ir_node *bl = get_nodes_block(irn);
342 int stack_dir = arch_env->stack_dir;
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;
367 /* Let the isa fill out the abi description for that call node. */
368 arch_env_get_call_abi(arch_env, call_tp, call);
370 /* Insert code to put the stack arguments on the stack. */
371 assert(get_Call_n_params(irn) == n_params);
372 stack_param_idx = ALLOCAN(int, n_params);
373 for (i = 0; i < n_params; ++i) {
374 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 0);
377 int arg_size = get_type_size_bytes(get_method_param_type(call_tp, i));
379 stack_size += round_up2(arg->space_before, arg->alignment);
380 stack_size += round_up2(arg_size, arg->alignment);
381 stack_size += round_up2(arg->space_after, arg->alignment);
383 stack_param_idx[n_stack_params++] = i;
387 /* Collect all arguments which are passed in registers. */
388 reg_param_idxs = ALLOCAN(int, n_params);
389 for (i = 0; i < n_params; ++i) {
390 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 0);
391 if (arg && arg->in_reg) {
392 reg_param_idxs[n_reg_params++] = i;
397 * If the stack is decreasing and we do not want to store sequentially,
398 * or someone else allocated the call frame
399 * we allocate as much space on the stack all parameters need, by
400 * moving the stack pointer along the stack's direction.
402 * Note: we also have to do this for stack_size == 0, because we may have
403 * to adjust stack alignment for the call.
405 if (stack_dir < 0 && !do_seq && !no_alloc) {
406 curr_sp = be_new_IncSP(sp, bl, curr_sp, stack_size, 1);
409 dbgi = get_irn_dbg_info(irn);
410 /* If there are some parameters which shall be passed on the stack. */
411 if (n_stack_params > 0) {
413 ir_node **in = ALLOCAN(ir_node*, n_stack_params+1);
417 * Reverse list of stack parameters if call arguments are from left to right.
418 * We must them reverse again if they are pushed (not stored) and the stack
419 * direction is downwards.
421 if (call->flags.bits.left_to_right ^ (do_seq && stack_dir < 0)) {
422 for (i = 0; i < n_stack_params >> 1; ++i) {
423 int other = n_stack_params - i - 1;
424 int tmp = stack_param_idx[i];
425 stack_param_idx[i] = stack_param_idx[other];
426 stack_param_idx[other] = tmp;
430 curr_mem = get_Call_mem(irn);
432 in[n_in++] = curr_mem;
435 for (i = 0; i < n_stack_params; ++i) {
436 int p = stack_param_idx[i];
437 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
438 ir_node *param = get_Call_param(irn, p);
439 ir_node *addr = curr_sp;
441 ir_type *param_type = get_method_param_type(call_tp, p);
442 int param_size = get_type_size_bytes(param_type) + arg->space_after;
445 * If we wanted to build the arguments sequentially,
446 * the stack pointer for the next must be incremented,
447 * and the memory value propagated.
451 addr = curr_sp = be_new_IncSP(sp, bl, curr_sp,
452 param_size + arg->space_before, 0);
453 add_irn_dep(curr_sp, curr_mem);
455 curr_ofs += arg->space_before;
456 curr_ofs = round_up2(curr_ofs, arg->alignment);
458 /* Make the expression to compute the argument's offset. */
460 ir_mode *constmode = mach_mode;
461 if (mode_is_reference(mach_mode)) {
464 addr = new_r_Const_long(irg, constmode, curr_ofs);
465 addr = new_r_Add(bl, curr_sp, addr, mach_mode);
469 /* Insert a store for primitive arguments. */
470 if (is_atomic_type(param_type)) {
472 ir_node *mem_input = do_seq ? curr_mem : new_r_NoMem(irg);
473 store = new_rd_Store(dbgi, bl, mem_input, addr, param, 0);
474 mem = new_r_Proj(store, mode_M, pn_Store_M);
476 /* Make a mem copy for compound arguments. */
479 assert(mode_is_reference(get_irn_mode(param)));
480 copy = new_rd_CopyB(dbgi, bl, curr_mem, addr, param, param_type);
481 mem = new_r_Proj(copy, mode_M, pn_CopyB_M);
484 curr_ofs += param_size;
492 /* We need the sync only, if we didn't build the stores sequentially. */
494 if (n_stack_params >= 1) {
495 curr_mem = new_r_Sync(bl, n_in, in);
497 curr_mem = get_Call_mem(irn);
502 /* check for the return_twice property */
503 destroy_all_regs = 0;
504 if (is_SymConst_addr_ent(call_ptr)) {
505 ir_entity *ent = get_SymConst_entity(call_ptr);
507 if (get_entity_additional_properties(ent) & mtp_property_returns_twice)
508 destroy_all_regs = 1;
510 ir_type *call_tp = get_Call_type(irn);
512 if (get_method_additional_properties(call_tp) & mtp_property_returns_twice)
513 destroy_all_regs = 1;
516 /* Put caller save into the destroyed set and state registers in the states
518 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
520 const arch_register_class_t *cls = &arch_env->register_classes[i];
521 for (j = 0; j < cls->n_regs; ++j) {
522 const arch_register_t *reg = arch_register_for_index(cls, j);
524 /* even if destroyed all is specified, neither SP nor FP are
525 * destroyed (else bad things will happen) */
526 if (reg == arch_env->sp || reg == arch_env->bp)
529 if (arch_register_type_is(reg, state)) {
530 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
531 ARR_APP1(const arch_register_t*, states, reg);
532 /* we're already in the destroyed set so no need for further
536 if (destroy_all_regs || arch_register_type_is(reg, caller_save)) {
537 if (! arch_register_type_is(reg, ignore)) {
538 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
544 /* search the largest result proj number */
545 res_projs = ALLOCANZ(ir_node*, n_res);
547 foreach_out_edge(irn, edge) {
548 const ir_edge_t *res_edge;
549 ir_node *irn = get_edge_src_irn(edge);
551 if (!is_Proj(irn) || get_Proj_proj(irn) != pn_Call_T_result)
554 foreach_out_edge(irn, res_edge) {
556 ir_node *res = get_edge_src_irn(res_edge);
558 assert(is_Proj(res));
560 proj = get_Proj_proj(res);
561 assert(proj < n_res);
562 assert(res_projs[proj] == NULL);
563 res_projs[proj] = res;
569 /** TODO: this is not correct for cases where return values are passed
570 * on the stack, but no known ABI does this currently...
572 n_reg_results = n_res;
575 in = ALLOCAN(ir_node*, n_reg_params + ARR_LEN(states));
577 /* make the back end call node and set its register requirements. */
578 for (i = 0; i < n_reg_params; ++i) {
579 in[n_ins++] = get_Call_param(irn, reg_param_idxs[i]);
582 /* add state registers ins */
583 for (i = 0; i < ARR_LEN(states); ++i) {
584 const arch_register_t *reg = states[i];
585 const arch_register_class_t *cls = arch_register_get_class(reg);
587 ir_node *regnode = be_abi_reg_map_get(env->regs, reg);
588 ir_fprintf(stderr, "Adding %+F\n", regnode);
590 ir_node *regnode = new_r_Unknown(irg, arch_register_class_mode(cls));
591 in[n_ins++] = regnode;
593 assert(n_ins == (int) (n_reg_params + ARR_LEN(states)));
595 /* ins collected, build the call */
596 if (env->call->flags.bits.call_has_imm && is_SymConst(call_ptr)) {
598 low_call = be_new_Call(dbgi, irg, bl, curr_mem, curr_sp, curr_sp,
599 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
600 n_ins, in, get_Call_type(irn));
601 be_Call_set_entity(low_call, get_SymConst_entity(call_ptr));
604 low_call = be_new_Call(dbgi, irg, bl, curr_mem, curr_sp, call_ptr,
605 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
606 n_ins, in, get_Call_type(irn));
608 be_Call_set_pop(low_call, call->pop);
610 /* put the call into the list of all calls for later processing */
611 ARR_APP1(ir_node *, env->calls, low_call);
613 /* create new stack pointer */
614 curr_sp = new_r_Proj(low_call, get_irn_mode(curr_sp), pn_be_Call_sp);
615 be_set_constr_single_reg_out(low_call, pn_be_Call_sp, sp,
616 arch_register_req_type_ignore | arch_register_req_type_produces_sp);
617 arch_set_irn_register(curr_sp, sp);
619 /* now handle results */
620 for (i = 0; i < n_res; ++i) {
622 ir_node *proj = res_projs[i];
623 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
625 /* returns values on stack not supported yet */
629 shift the proj number to the right, since we will drop the
630 unspeakable Proj_T from the Call. Therefore, all real argument
631 Proj numbers must be increased by pn_be_Call_first_res
633 pn = i + pn_be_Call_first_res;
636 ir_type *res_type = get_method_res_type(call_tp, i);
637 ir_mode *mode = get_type_mode(res_type);
638 proj = new_r_Proj(low_call, mode, pn);
641 set_Proj_pred(proj, low_call);
642 set_Proj_proj(proj, pn);
646 /* remove register from destroyed regs */
648 int n = ARR_LEN(destroyed_regs);
649 for (j = 0; j < n; ++j) {
650 if (destroyed_regs[j] == arg->reg) {
651 destroyed_regs[j] = destroyed_regs[n-1];
652 ARR_SHRINKLEN(destroyed_regs,n-1);
660 Set the register class of the call address to
661 the backend provided class (default: stack pointer class)
663 be_node_set_reg_class_in(low_call, be_pos_Call_ptr, call->cls_addr);
665 DBG((dbg, LEVEL_3, "\tcreated backend call %+F\n", low_call));
667 /* Set the register classes and constraints of the Call parameters. */
668 for (i = 0; i < n_reg_params; ++i) {
669 int index = reg_param_idxs[i];
670 be_abi_call_arg_t *arg = get_call_arg(call, 0, index, 0);
671 assert(arg->reg != NULL);
673 be_set_constr_single_reg_in(low_call, be_pos_Call_first_arg + i,
677 /* Set the register constraints of the results. */
678 for (i = 0; i < n_res; ++i) {
679 ir_node *proj = res_projs[i];
680 const be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
681 int pn = get_Proj_proj(proj);
684 be_set_constr_single_reg_out(low_call, pn, arg->reg, 0);
685 arch_set_irn_register(proj, arg->reg);
687 exchange(irn, low_call);
689 /* kill the ProjT node */
690 if (res_proj != NULL) {
694 /* Make additional projs for the caller save registers
695 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 (i = 0; i < ARR_LEN(destroyed_regs); ++i) {
711 const arch_register_t *reg = destroyed_regs[i];
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, 0);
716 arch_set_irn_register(proj, reg);
718 set_irn_link(proj, (void*) reg);
723 for (i = 0; i < n_reg_results; ++i) {
724 ir_node *proj = res_projs[i];
725 const arch_register_t *reg = arch_get_irn_register(proj);
726 set_irn_link(proj, (void*) reg);
731 /* create the Keep for the caller save registers */
732 keep = be_new_Keep(bl, n, in);
733 for (i = 0; i < n; ++i) {
734 const arch_register_t *reg = get_irn_link(in[i]);
735 be_node_set_reg_class_in(keep, i, reg->reg_class);
739 /* Clean up the stack. */
740 assert(stack_size >= call->pop);
741 stack_size -= call->pop;
743 if (stack_size > 0) {
744 ir_node *mem_proj = NULL;
746 foreach_out_edge(low_call, edge) {
747 ir_node *irn = get_edge_src_irn(edge);
748 if (is_Proj(irn) && get_Proj_proj(irn) == pn_Call_M) {
755 mem_proj = new_r_Proj(low_call, mode_M, pn_be_Call_M_regular);
756 keep_alive(mem_proj);
759 /* Clean up the stack frame or revert alignment fixes if we allocated it */
761 curr_sp = be_new_IncSP(sp, bl, curr_sp, -stack_size, 0);
764 be_abi_call_free(call);
767 DEL_ARR_F(destroyed_regs);
773 * Adjust the size of a node representing a stack alloc or free for the minimum stack alignment.
775 * @param alignment the minimum stack alignment
776 * @param size the node containing the non-aligned size
777 * @param block the block where new nodes are allocated on
778 * @param dbg debug info for new nodes
780 * @return a node representing the aligned size
782 static ir_node *adjust_alloc_size(unsigned stack_alignment, ir_node *size,
783 ir_node *block, dbg_info *dbg)
785 if (stack_alignment > 1) {
791 assert(is_po2(stack_alignment));
793 mode = get_irn_mode(size);
794 tv = new_tarval_from_long(stack_alignment-1, mode);
795 irg = get_Block_irg(block);
796 mask = new_r_Const(irg, tv);
797 size = new_rd_Add(dbg, block, size, mask, mode);
799 tv = new_tarval_from_long(-(long)stack_alignment, mode);
800 mask = new_r_Const(irg, tv);
801 size = new_rd_And(dbg, block, size, mask, mode);
807 * The alloca is transformed into a back end alloca node and connected to the stack nodes.
809 static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp)
811 ir_node *block = get_nodes_block(alloc);
812 ir_graph *irg = get_Block_irg(block);
813 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
814 ir_node *alloc_mem = NULL;
815 ir_node *alloc_res = NULL;
816 ir_type *type = get_Alloc_type(alloc);
819 const ir_edge_t *edge;
824 unsigned stack_alignment;
826 /* all non-stack Alloc nodes should already be lowered before the backend */
827 assert(get_Alloc_where(alloc) == stack_alloc);
829 foreach_out_edge(alloc, edge) {
830 ir_node *irn = get_edge_src_irn(edge);
832 assert(is_Proj(irn));
833 switch (get_Proj_proj(irn)) {
845 /* Beware: currently Alloc nodes without a result might happen,
846 only escape analysis kills them and this phase runs only for object
847 oriented source. We kill the Alloc here. */
848 if (alloc_res == NULL && alloc_mem) {
849 exchange(alloc_mem, get_Alloc_mem(alloc));
853 dbg = get_irn_dbg_info(alloc);
854 count = get_Alloc_count(alloc);
856 /* we might need to multiply the count with the element size */
857 if (type != firm_unknown_type && get_type_size_bytes(type) != 1) {
858 ir_mode *mode = get_irn_mode(count);
859 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type),
861 ir_node *cnst = new_rd_Const(dbg, irg, tv);
862 size = new_rd_Mul(dbg, block, count, cnst, mode);
867 /* The stack pointer will be modified in an unknown manner.
868 We cannot omit it. */
869 env->call->flags.bits.try_omit_fp = 0;
871 stack_alignment = 1 << arch_env->stack_alignment;
872 size = adjust_alloc_size(stack_alignment, size, block, dbg);
873 new_alloc = be_new_AddSP(arch_env->sp, block, curr_sp, size);
874 set_irn_dbg_info(new_alloc, dbg);
876 if (alloc_mem != NULL) {
880 addsp_mem = new_r_Proj(new_alloc, mode_M, pn_be_AddSP_M);
882 /* We need to sync the output mem of the AddSP with the input mem
883 edge into the alloc node. */
884 ins[0] = get_Alloc_mem(alloc);
886 sync = new_r_Sync(block, 2, ins);
888 exchange(alloc_mem, sync);
891 exchange(alloc, new_alloc);
893 /* fix projnum of alloca res */
894 set_Proj_proj(alloc_res, pn_be_AddSP_res);
896 curr_sp = new_r_Proj(new_alloc, get_irn_mode(curr_sp), pn_be_AddSP_sp);
903 * The Free is transformed into a back end free node and connected to the stack nodes.
905 static ir_node *adjust_free(be_abi_irg_t *env, ir_node *free, ir_node *curr_sp)
907 ir_node *block = get_nodes_block(free);
908 ir_graph *irg = get_irn_irg(free);
909 ir_type *type = get_Free_type(free);
910 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
911 ir_mode *sp_mode = arch_env->sp->reg_class->mode;
912 dbg_info *dbg = get_irn_dbg_info(free);
913 ir_node *subsp, *mem, *res, *size, *sync;
915 unsigned stack_alignment;
917 /* all non-stack-alloc Free nodes should already be lowered before the
919 assert(get_Free_where(free) == stack_alloc);
921 /* we might need to multiply the size with the element size */
922 if (type != firm_unknown_type && get_type_size_bytes(type) != 1) {
923 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type), mode_Iu);
924 ir_node *cnst = new_rd_Const(dbg, irg, tv);
925 ir_node *mul = new_rd_Mul(dbg, block, get_Free_size(free),
929 size = get_Free_size(free);
932 stack_alignment = 1 << arch_env->stack_alignment;
933 size = adjust_alloc_size(stack_alignment, size, block, dbg);
935 /* The stack pointer will be modified in an unknown manner.
936 We cannot omit it. */
937 env->call->flags.bits.try_omit_fp = 0;
938 subsp = be_new_SubSP(arch_env->sp, block, curr_sp, size);
939 set_irn_dbg_info(subsp, dbg);
941 mem = new_r_Proj(subsp, mode_M, pn_be_SubSP_M);
942 res = new_r_Proj(subsp, sp_mode, pn_be_SubSP_sp);
944 /* we need to sync the memory */
945 in[0] = get_Free_mem(free);
947 sync = new_r_Sync(block, 2, in);
949 /* and make the AddSP dependent on the former memory */
950 add_irn_dep(subsp, get_Free_mem(free));
953 exchange(free, sync);
960 * Check if a node is somehow data dependent on another one.
961 * both nodes must be in the same basic block.
962 * @param n1 The first node.
963 * @param n2 The second node.
964 * @return 1, if n1 is data dependent (transitively) on n2, 0 if not.
966 static int dependent_on(ir_node *n1, ir_node *n2)
968 assert(get_nodes_block(n1) == get_nodes_block(n2));
970 return heights_reachable_in_block(ir_heights, n1, n2);
973 static int cmp_call_dependency(const void *c1, const void *c2)
975 ir_node *n1 = *(ir_node **) c1;
976 ir_node *n2 = *(ir_node **) c2;
979 Classical qsort() comparison function behavior:
980 0 if both elements are equal
981 1 if second is "smaller" that first
982 -1 if first is "smaller" that second
984 if (dependent_on(n1, n2))
987 if (dependent_on(n2, n1))
990 /* The nodes have no depth order, but we need a total order because qsort()
992 return get_irn_idx(n1) - get_irn_idx(n2);
996 * Walker: links all Call/Alloc/Free nodes to the Block they are contained.
997 * Clears the irg_is_leaf flag if a Call is detected.
999 static void link_ops_in_block_walker(ir_node *irn, void *data)
1001 be_abi_irg_t *env = data;
1002 ir_opcode code = get_irn_opcode(irn);
1004 if (code == iro_Call ||
1005 (code == iro_Alloc && get_Alloc_where(irn) == stack_alloc) ||
1006 (code == iro_Free && get_Free_where(irn) == stack_alloc)) {
1007 ir_node *bl = get_nodes_block(irn);
1008 void *save = get_irn_link(bl);
1010 if (code == iro_Call)
1011 env->call->flags.bits.irg_is_leaf = 0;
1013 set_irn_link(irn, save);
1014 set_irn_link(bl, irn);
1017 if (code == iro_Builtin && get_Builtin_kind(irn) == ir_bk_return_address) {
1018 ir_node *param = get_Builtin_param(irn, 0);
1019 ir_tarval *tv = get_Const_tarval(param);
1020 unsigned long value = get_tarval_long(tv);
1021 /* use ebp, so the climbframe algo works... */
1023 env->call->flags.bits.try_omit_fp = 0;
1030 * Process all Call/Alloc/Free nodes inside a basic block.
1031 * Note that the link field of the block must contain a linked list of all
1032 * Call nodes inside the Block. We first order this list according to data dependency
1033 * and that connect the calls together.
1035 static void process_ops_in_block(ir_node *bl, void *data)
1037 be_abi_irg_t *env = data;
1038 ir_node *curr_sp = env->init_sp;
1045 for (irn = get_irn_link(bl); irn != NULL; irn = get_irn_link(irn)) {
1049 nodes = ALLOCAN(ir_node*, n_nodes);
1050 for (irn = get_irn_link(bl), n = 0; irn; irn = get_irn_link(irn), ++n) {
1054 /* If there were call nodes in the block. */
1059 /* order the call nodes according to data dependency */
1060 qsort(nodes, n_nodes, sizeof(nodes[0]), cmp_call_dependency);
1062 for (i = n_nodes - 1; i >= 0; --i) {
1063 ir_node *irn = nodes[i];
1065 DBG((dbg, LEVEL_3, "\tprocessing call %+F\n", irn));
1066 switch (get_irn_opcode(irn)) {
1069 /* The stack pointer will be modified due to a call. */
1070 env->call->flags.bits.try_omit_fp = 0;
1072 curr_sp = adjust_call(env, irn, curr_sp);
1075 if (get_Alloc_where(irn) == stack_alloc)
1076 curr_sp = adjust_alloc(env, irn, curr_sp);
1079 if (get_Free_where(irn) == stack_alloc)
1080 curr_sp = adjust_free(env, irn, curr_sp);
1083 panic("invalid call");
1087 /* Keep the last stack state in the block by tying it to Keep node,
1088 * the proj from calls is already kept */
1089 if (curr_sp != env->init_sp &&
1090 !(is_Proj(curr_sp) && be_is_Call(get_Proj_pred(curr_sp)))) {
1092 keep = be_new_Keep(bl, 1, nodes);
1093 pmap_insert(env->keep_map, bl, keep);
1097 set_irn_link(bl, curr_sp);
1101 * Adjust all call nodes in the graph to the ABI conventions.
1103 static void process_calls(ir_graph *irg)
1105 be_abi_irg_t *abi = be_get_irg_abi(irg);
1107 abi->call->flags.bits.irg_is_leaf = 1;
1108 irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, abi);
1110 ir_heights = heights_new(irg);
1111 irg_block_walk_graph(irg, NULL, process_ops_in_block, abi);
1112 heights_free(ir_heights);
1116 * Computes the stack argument layout type.
1117 * Changes a possibly allocated value param type by moving
1118 * entities to the stack layout type.
1120 * @param env the ABI environment
1121 * @param call the current call ABI
1122 * @param method_type the method type
1123 * @param val_param_tp the value parameter type, will be destroyed
1124 * @param param_map an array mapping method arguments to the stack layout type
1126 * @return the stack argument layout type
1128 static ir_type *compute_arg_type(be_abi_irg_t *env, ir_graph *irg,
1129 be_abi_call_t *call,
1130 ir_type *method_type, ir_type *val_param_tp,
1131 ir_entity ***param_map)
1133 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1134 int dir = env->call->flags.bits.left_to_right ? 1 : -1;
1135 int inc = arch_env->stack_dir * dir;
1136 int n = get_method_n_params(method_type);
1137 int curr = inc > 0 ? 0 : n - 1;
1138 struct obstack *obst = be_get_be_obst(irg);
1144 ident *id = get_entity_ident(get_irg_entity(irg));
1147 *param_map = map = OALLOCN(obst, ir_entity*, n);
1148 res = new_type_struct(id_mangle_u(id, new_id_from_chars("arg_type", 8)));
1149 for (i = 0; i < n; ++i, curr += inc) {
1150 ir_type *param_type = get_method_param_type(method_type, curr);
1151 be_abi_call_arg_t *arg = get_call_arg(call, 0, curr, 1);
1154 if (arg->on_stack) {
1155 if (val_param_tp != NULL) {
1156 /* the entity was already created, create a copy in the param type */
1157 ir_entity *val_ent = get_method_value_param_ent(method_type, i);
1158 arg->stack_ent = copy_entity_own(val_ent, res);
1159 set_entity_link(val_ent, arg->stack_ent);
1160 set_entity_link(arg->stack_ent, NULL);
1162 /* create a new entity */
1163 snprintf(buf, sizeof(buf), "param_%d", i);
1164 arg->stack_ent = new_entity(res, new_id_from_str(buf), param_type);
1166 ofs += arg->space_before;
1167 ofs = round_up2(ofs, arg->alignment);
1168 set_entity_offset(arg->stack_ent, ofs);
1169 ofs += arg->space_after;
1170 ofs += get_type_size_bytes(param_type);
1171 map[i] = arg->stack_ent;
1174 set_type_size_bytes(res, ofs);
1175 set_type_state(res, layout_fixed);
1180 const arch_register_t *reg;
1184 static int cmp_regs(const void *a, const void *b)
1186 const reg_node_map_t *p = a;
1187 const reg_node_map_t *q = b;
1189 if (p->reg->reg_class == q->reg->reg_class)
1190 return p->reg->index - q->reg->index;
1192 return p->reg->reg_class - q->reg->reg_class;
1195 static void reg_map_to_arr(reg_node_map_t *res, pmap *reg_map)
1198 int n = pmap_count(reg_map);
1201 foreach_pmap(reg_map, ent) {
1202 res[i].reg = ent->key;
1203 res[i].irn = ent->value;
1207 qsort(res, n, sizeof(res[0]), cmp_regs);
1211 * Creates a barrier.
1213 static ir_node *create_barrier(ir_node *bl, ir_node **mem, pmap *regs,
1216 int n_regs = pmap_count(regs);
1222 in = ALLOCAN(ir_node*, n_regs+1);
1223 rm = ALLOCAN(reg_node_map_t, n_regs);
1224 reg_map_to_arr(rm, regs);
1225 for (n = 0; n < n_regs; ++n) {
1233 irn = be_new_Barrier(bl, n, in);
1235 for (n = 0; n < n_regs; ++n) {
1236 ir_node *pred = rm[n].irn;
1237 const arch_register_t *reg = rm[n].reg;
1238 arch_register_type_t add_type = 0;
1240 const backend_info_t *info;
1242 /* stupid workaround for now... as not all nodes report register
1244 info = be_get_info(skip_Proj(pred));
1245 if (info != NULL && info->out_infos != NULL) {
1246 const arch_register_req_t *ireq = arch_get_register_req_out(pred);
1247 if (ireq->type & arch_register_req_type_ignore)
1248 add_type |= arch_register_req_type_ignore;
1249 if (ireq->type & arch_register_req_type_produces_sp)
1250 add_type |= arch_register_req_type_produces_sp;
1253 proj = new_r_Proj(irn, get_irn_mode(pred), n);
1254 be_node_set_reg_class_in(irn, n, reg->reg_class);
1256 be_set_constr_single_reg_in(irn, n, reg, 0);
1257 be_set_constr_single_reg_out(irn, n, reg, add_type);
1258 arch_set_irn_register(proj, reg);
1260 pmap_insert(regs, (void *) reg, proj);
1264 *mem = new_r_Proj(irn, mode_M, n);
1271 * Creates a be_Return for a Return node.
1273 * @param @env the abi environment
1274 * @param irn the Return node or NULL if there was none
1275 * @param bl the block where the be_Retun should be placed
1276 * @param mem the current memory
1277 * @param n_res number of return results
1279 static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl,
1280 ir_node *mem, int n_res)
1282 be_abi_call_t *call = env->call;
1283 ir_graph *irg = get_Block_irg(bl);
1284 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1286 pmap *reg_map = pmap_create();
1287 ir_node *keep = pmap_get(env->keep_map, bl);
1294 const arch_register_t **regs;
1298 get the valid stack node in this block.
1299 If we had a call in that block there is a Keep constructed by process_calls()
1300 which points to the last stack modification in that block. we'll use
1301 it then. Else we use the stack from the start block and let
1302 the ssa construction fix the usage.
1304 stack = be_abi_reg_map_get(env->regs, arch_env->sp);
1306 stack = get_irn_n(keep, 0);
1308 remove_End_keepalive(get_irg_end(irg), keep);
1311 /* Insert results for Return into the register map. */
1312 for (i = 0; i < n_res; ++i) {
1313 ir_node *res = get_Return_res(irn, i);
1314 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1315 assert(arg->in_reg && "return value must be passed in register");
1316 pmap_insert(reg_map, (void *) arg->reg, res);
1319 /* Add uses of the callee save registers. */
1320 foreach_pmap(env->regs, ent) {
1321 const arch_register_t *reg = ent->key;
1322 if (arch_register_type_is(reg, callee_save) || arch_register_type_is(reg, ignore))
1323 pmap_insert(reg_map, ent->key, ent->value);
1326 be_abi_reg_map_set(reg_map, arch_env->sp, stack);
1328 /* Make the Epilogue node and call the arch's epilogue maker. */
1329 create_barrier(bl, &mem, reg_map, 1);
1330 call->cb->epilogue(env->cb, bl, &mem, reg_map);
1333 Maximum size of the in array for Return nodes is
1334 return args + callee save/ignore registers + memory + stack pointer
1336 in_max = pmap_count(reg_map) + n_res + 2;
1338 in = ALLOCAN(ir_node*, in_max);
1339 regs = ALLOCAN(arch_register_t const*, in_max);
1342 in[1] = be_abi_reg_map_get(reg_map, arch_env->sp);
1344 regs[1] = arch_env->sp;
1347 /* clear SP entry, since it has already been grown. */
1348 pmap_insert(reg_map, (void *) arch_env->sp, NULL);
1349 for (i = 0; i < n_res; ++i) {
1350 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1352 in[n] = be_abi_reg_map_get(reg_map, arg->reg);
1353 regs[n++] = arg->reg;
1355 /* Clear the map entry to mark the register as processed. */
1356 be_abi_reg_map_set(reg_map, arg->reg, NULL);
1359 /* grow the rest of the stuff. */
1360 foreach_pmap(reg_map, ent) {
1363 regs[n++] = ent->key;
1367 /* The in array for the new back end return is now ready. */
1369 dbgi = get_irn_dbg_info(irn);
1373 /* we have to pop the shadow parameter in in case of struct returns */
1375 ret = be_new_Return(dbgi, irg, bl, n_res, pop, n, in);
1377 /* Set the register classes of the return's parameter accordingly. */
1378 for (i = 0; i < n; ++i) {
1379 if (regs[i] == NULL)
1382 be_node_set_reg_class_in(ret, i, regs[i]->reg_class);
1385 /* Free the space of the Epilog's in array and the register <-> proj map. */
1386 pmap_destroy(reg_map);
1391 typedef struct ent_pos_pair ent_pos_pair;
1392 struct ent_pos_pair {
1393 ir_entity *ent; /**< a value param entity */
1394 int pos; /**< its parameter number */
1395 ent_pos_pair *next; /**< for linking */
1398 typedef struct lower_frame_sels_env_t {
1399 ent_pos_pair *value_param_list; /**< the list of all value param entities */
1400 ir_node *frame; /**< the current frame */
1401 const arch_register_class_t *sp_class; /**< register class of the stack pointer */
1402 const arch_register_class_t *link_class; /**< register class of the link pointer */
1403 ir_type *value_tp; /**< the value type if any */
1404 ir_type *frame_tp; /**< the frame type */
1405 int static_link_pos; /**< argument number of the hidden static link */
1406 } lower_frame_sels_env_t;
1409 * Return an entity from the backend for an value param entity.
1411 * @param ent an value param type entity
1412 * @param ctx context
1414 static ir_entity *get_argument_entity(ir_entity *ent, lower_frame_sels_env_t *ctx)
1416 ir_entity *argument_ent = get_entity_link(ent);
1418 if (argument_ent == NULL) {
1419 /* we have NO argument entity yet: This is bad, as we will
1420 * need one for backing store.
1423 ir_type *frame_tp = ctx->frame_tp;
1424 unsigned offset = get_type_size_bytes(frame_tp);
1425 ir_type *tp = get_entity_type(ent);
1426 unsigned align = get_type_alignment_bytes(tp);
1428 offset += align - 1;
1429 offset &= ~(align - 1);
1431 argument_ent = copy_entity_own(ent, frame_tp);
1433 /* must be automatic to set a fixed layout */
1434 set_entity_offset(argument_ent, offset);
1435 offset += get_type_size_bytes(tp);
1437 set_type_size_bytes(frame_tp, offset);
1438 set_entity_link(ent, argument_ent);
1440 return argument_ent;
1443 * Walker: Replaces Sels of frame type and
1444 * value param type entities by FrameAddress.
1445 * Links all used entities.
1447 static void lower_frame_sels_walker(ir_node *irn, void *data)
1449 lower_frame_sels_env_t *ctx = data;
1452 ir_node *ptr = get_Sel_ptr(irn);
1454 if (ptr == ctx->frame) {
1455 ir_entity *ent = get_Sel_entity(irn);
1456 ir_node *bl = get_nodes_block(irn);
1459 int is_value_param = 0;
1461 if (get_entity_owner(ent) == ctx->value_tp) {
1464 /* replace by its copy from the argument type */
1465 pos = get_struct_member_index(ctx->value_tp, ent);
1466 ent = get_argument_entity(ent, ctx);
1469 nw = be_new_FrameAddr(ctx->sp_class, bl, ctx->frame, ent);
1472 /* check, if it's a param Sel and if have not seen this entity before */
1473 if (is_value_param && get_entity_link(ent) == NULL) {
1479 ARR_APP1(ent_pos_pair, ctx->value_param_list, pair);
1481 set_entity_link(ent, ctx->value_param_list);
1488 * Check if a value parameter is transmitted as a register.
1489 * This might happen if the address of an parameter is taken which is
1490 * transmitted in registers.
1492 * Note that on some architectures this case must be handled specially
1493 * because the place of the backing store is determined by their ABI.
1495 * In the default case we move the entity to the frame type and create
1496 * a backing store into the first block.
1498 static void fix_address_of_parameter_access(be_abi_irg_t *env, ir_graph *irg,
1499 ent_pos_pair *value_param_list)
1501 be_abi_call_t *call = env->call;
1502 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1503 ent_pos_pair *entry, *new_list;
1505 int i, n = ARR_LEN(value_param_list);
1508 for (i = 0; i < n; ++i) {
1509 int pos = value_param_list[i].pos;
1510 be_abi_call_arg_t *arg = get_call_arg(call, 0, pos, 1);
1513 DBG((dbg, LEVEL_2, "\targ #%d need backing store\n", pos));
1514 value_param_list[i].next = new_list;
1515 new_list = &value_param_list[i];
1518 if (new_list != NULL) {
1519 /* ok, change the graph */
1520 ir_node *start_bl = get_irg_start_block(irg);
1521 ir_node *first_bl = get_first_block_succ(start_bl);
1522 ir_node *frame, *imem, *nmem, *store, *mem, *args;
1523 optimization_state_t state;
1526 assert(first_bl && first_bl != start_bl);
1527 /* we had already removed critical edges, so the following
1528 assertion should be always true. */
1529 assert(get_Block_n_cfgpreds(first_bl) == 1);
1531 /* now create backing stores */
1532 frame = get_irg_frame(irg);
1533 imem = get_irg_initial_mem(irg);
1535 save_optimization_state(&state);
1537 nmem = new_r_Proj(get_irg_start(irg), mode_M, pn_Start_M);
1538 restore_optimization_state(&state);
1540 /* reroute all edges to the new memory source */
1541 edges_reroute(imem, nmem, irg);
1545 args = get_irg_args(irg);
1546 for (entry = new_list; entry != NULL; entry = entry->next) {
1548 ir_type *tp = get_entity_type(entry->ent);
1549 ir_mode *mode = get_type_mode(tp);
1552 /* address for the backing store */
1553 addr = be_new_FrameAddr(arch_env->sp->reg_class, first_bl, frame, entry->ent);
1556 mem = new_r_Proj(store, mode_M, pn_Store_M);
1558 /* the backing store itself */
1559 store = new_r_Store(first_bl, mem, addr,
1560 new_r_Proj(args, mode, i), 0);
1562 /* the new memory Proj gets the last Proj from store */
1563 set_Proj_pred(nmem, store);
1564 set_Proj_proj(nmem, pn_Store_M);
1566 /* move all entities to the frame type */
1567 frame_tp = get_irg_frame_type(irg);
1568 offset = get_type_size_bytes(frame_tp);
1570 /* we will add new entities: set the layout to undefined */
1571 assert(get_type_state(frame_tp) == layout_fixed);
1572 set_type_state(frame_tp, layout_undefined);
1573 for (entry = new_list; entry != NULL; entry = entry->next) {
1574 ir_entity *ent = entry->ent;
1576 /* If the entity is still on the argument type, move it to the
1578 * This happens if the value_param type was build due to compound
1580 if (get_entity_owner(ent) != frame_tp) {
1581 ir_type *tp = get_entity_type(ent);
1582 unsigned align = get_type_alignment_bytes(tp);
1584 offset += align - 1;
1585 offset &= ~(align - 1);
1586 set_entity_owner(ent, frame_tp);
1587 /* must be automatic to set a fixed layout */
1588 set_entity_offset(ent, offset);
1589 offset += get_type_size_bytes(tp);
1592 set_type_size_bytes(frame_tp, offset);
1593 /* fix the layout again */
1594 set_type_state(frame_tp, layout_fixed);
1599 * The start block has no jump, instead it has an initial exec Proj.
1600 * The backend wants to handle all blocks the same way, so we replace
1601 * the out cfg edge with a real jump.
1603 static void fix_start_block(ir_graph *irg)
1605 ir_node *initial_X = get_irg_initial_exec(irg);
1606 ir_node *start_block = get_irg_start_block(irg);
1607 const ir_edge_t *edge;
1609 assert(is_Proj(initial_X));
1611 foreach_out_edge(initial_X, edge) {
1612 ir_node *block = get_edge_src_irn(edge);
1614 if (is_Anchor(block))
1616 if (block != start_block) {
1617 ir_node *jmp = new_r_Jmp(start_block);
1618 set_Block_cfgpred(block, get_edge_src_pos(edge), jmp);
1619 set_irg_initial_exec(irg, jmp);
1623 panic("Initial exec has no follow block in %+F", irg);
1627 * Update the entity of Sels to the outer value parameters.
1629 static void update_outer_frame_sels(ir_node *irn, void *env)
1631 lower_frame_sels_env_t *ctx = env;
1638 ptr = get_Sel_ptr(irn);
1639 if (! is_arg_Proj(ptr))
1641 if (get_Proj_proj(ptr) != ctx->static_link_pos)
1643 ent = get_Sel_entity(irn);
1645 if (get_entity_owner(ent) == ctx->value_tp) {
1646 /* replace by its copy from the argument type */
1647 pos = get_struct_member_index(ctx->value_tp, ent);
1648 ent = get_argument_entity(ent, ctx);
1649 set_Sel_entity(irn, ent);
1651 /* check, if we have not seen this entity before */
1652 if (get_entity_link(ent) == NULL) {
1658 ARR_APP1(ent_pos_pair, ctx->value_param_list, pair);
1660 set_entity_link(ent, ctx->value_param_list);
1666 * Fix access to outer local variables.
1668 static void fix_outer_variable_access(be_abi_irg_t *env,
1669 lower_frame_sels_env_t *ctx)
1675 for (i = get_class_n_members(ctx->frame_tp) - 1; i >= 0; --i) {
1676 ir_entity *ent = get_class_member(ctx->frame_tp, i);
1678 if (! is_method_entity(ent))
1681 irg = get_entity_irg(ent);
1686 * FIXME: find the number of the static link parameter
1687 * for now we assume 0 here
1689 ctx->static_link_pos = 0;
1691 irg_walk_graph(irg, NULL, update_outer_frame_sels, ctx);
1696 * Modify the irg itself and the frame type.
1698 static void modify_irg(ir_graph *irg)
1700 be_abi_irg_t *env = be_get_irg_abi(irg);
1701 be_abi_call_t *call = env->call;
1702 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1703 const arch_register_t *sp = arch_env->sp;
1704 ir_type *method_type = get_entity_type(get_irg_entity(irg));
1705 be_irg_t *birg = be_birg_from_irg(irg);
1706 struct obstack *obst = be_get_be_obst(irg);
1707 be_stack_layout_t *stack_layout = be_get_irg_stack_layout(irg);
1710 ir_node *new_mem_proj;
1716 unsigned frame_size;
1719 const arch_register_t *fp_reg;
1720 ir_node *frame_pointer;
1724 const ir_edge_t *edge;
1725 ir_type *arg_type, *bet_type, *tp;
1726 lower_frame_sels_env_t ctx;
1727 ir_entity **param_map;
1729 DBG((dbg, LEVEL_1, "introducing abi on %+F\n", irg));
1731 /* Must fetch memory here, otherwise the start Barrier gets the wrong
1732 * memory, which leads to loops in the DAG. */
1733 old_mem = get_irg_initial_mem(irg);
1735 irp_reserve_resources(irp, IR_RESOURCE_ENTITY_LINK);
1737 /* set the links of all frame entities to NULL, we use it
1738 to detect if an entity is already linked in the value_param_list */
1739 tp = get_method_value_param_type(method_type);
1742 /* clear the links of the clone type, let the
1743 original entities point to its clones */
1744 for (i = get_struct_n_members(tp) - 1; i >= 0; --i) {
1745 ir_entity *mem = get_struct_member(tp, i);
1746 set_entity_link(mem, NULL);
1750 arg_type = compute_arg_type(env, irg, call, method_type, tp, ¶m_map);
1752 /* Convert the Sel nodes in the irg to frame addr nodes: */
1753 ctx.value_param_list = NEW_ARR_F(ent_pos_pair, 0);
1754 ctx.frame = get_irg_frame(irg);
1755 ctx.sp_class = arch_env->sp->reg_class;
1756 ctx.link_class = arch_env->link_class;
1757 ctx.frame_tp = get_irg_frame_type(irg);
1759 /* layout the stackframe now */
1760 if (get_type_state(ctx.frame_tp) == layout_undefined) {
1761 default_layout_compound_type(ctx.frame_tp);
1764 /* we will possible add new entities to the frame: set the layout to undefined */
1765 assert(get_type_state(ctx.frame_tp) == layout_fixed);
1766 set_type_state(ctx.frame_tp, layout_undefined);
1768 irg_walk_graph(irg, lower_frame_sels_walker, NULL, &ctx);
1770 /* fix the frame type layout again */
1771 set_type_state(ctx.frame_tp, layout_fixed);
1772 /* align stackframe to 4 byte */
1773 frame_size = get_type_size_bytes(ctx.frame_tp);
1774 if (frame_size % 4 != 0) {
1775 set_type_size_bytes(ctx.frame_tp, frame_size + 4 - (frame_size % 4));
1778 env->regs = pmap_create();
1780 n_params = get_method_n_params(method_type);
1781 args = OALLOCNZ(obst, ir_node*, n_params);
1784 * for inner function we must now fix access to outer frame entities.
1786 fix_outer_variable_access(env, &ctx);
1788 /* Check if a value parameter is transmitted as a register.
1789 * This might happen if the address of an parameter is taken which is
1790 * transmitted in registers.
1792 * Note that on some architectures this case must be handled specially
1793 * because the place of the backing store is determined by their ABI.
1795 * In the default case we move the entity to the frame type and create
1796 * a backing store into the first block.
1798 fix_address_of_parameter_access(env, irg, ctx.value_param_list);
1800 DEL_ARR_F(ctx.value_param_list);
1801 irp_free_resources(irp, IR_RESOURCE_ENTITY_LINK);
1803 /* Fill the argument vector */
1804 arg_tuple = get_irg_args(irg);
1805 foreach_out_edge(arg_tuple, edge) {
1806 ir_node *irn = get_edge_src_irn(edge);
1807 if (! is_Anchor(irn)) {
1808 int nr = get_Proj_proj(irn);
1810 DBG((dbg, LEVEL_2, "\treading arg: %d -> %+F\n", nr, irn));
1814 bet_type = call->cb->get_between_type(env->cb);
1815 stack_frame_init(stack_layout, arg_type, bet_type,
1816 get_irg_frame_type(irg), arch_env->stack_dir, param_map);
1817 stack_layout->sp_relative = call->flags.bits.try_omit_fp;
1819 /* Count the register params and add them to the number of Projs for the RegParams node */
1820 for (i = 0; i < n_params; ++i) {
1821 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1822 if (arg->in_reg && args[i]) {
1823 assert(arg->reg != sp && "cannot use stack pointer as parameter register");
1824 assert(i == get_Proj_proj(args[i]));
1826 /* For now, associate the register with the old Proj from Start representing that argument. */
1827 pmap_insert(env->regs, (void *) arg->reg, args[i]);
1828 DBG((dbg, LEVEL_2, "\targ #%d -> reg %s\n", i, arg->reg->name));
1832 /* Collect all callee-save registers */
1833 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
1834 const arch_register_class_t *cls = &arch_env->register_classes[i];
1835 for (j = 0; j < cls->n_regs; ++j) {
1836 const arch_register_t *reg = &cls->regs[j];
1837 if (arch_register_type_is(reg, callee_save) ||
1838 arch_register_type_is(reg, state)) {
1839 pmap_insert(env->regs, (void *) reg, NULL);
1844 /* handle start block here (place a jump in the block) */
1845 fix_start_block(irg);
1847 pmap_insert(env->regs, (void *) sp, NULL);
1848 pmap_insert(env->regs, (void *) arch_env->bp, NULL);
1849 start_bl = get_irg_start_block(irg);
1850 env->start = be_new_Start(NULL, start_bl, pmap_count(env->regs) + 1);
1853 * make proj nodes for the callee save registers.
1854 * memorize them, since Return nodes get those as inputs.
1856 * Note, that if a register corresponds to an argument, the regs map contains
1857 * the old Proj from start for that argument.
1860 rm = ALLOCAN(reg_node_map_t, pmap_count(env->regs));
1861 reg_map_to_arr(rm, env->regs);
1862 for (i = 0, n = pmap_count(env->regs); i < n; ++i) {
1863 arch_register_t *reg = (void *) rm[i].reg;
1864 ir_mode *mode = reg->reg_class->mode;
1866 arch_register_req_type_t add_type = 0;
1870 add_type |= arch_register_req_type_produces_sp | arch_register_req_type_ignore;
1873 proj = new_r_Proj(env->start, mode, nr + 1);
1874 pmap_insert(env->regs, (void *) reg, proj);
1875 be_set_constr_single_reg_out(env->start, nr + 1, reg, add_type);
1876 arch_set_irn_register(proj, reg);
1878 DBG((dbg, LEVEL_2, "\tregister save proj #%d -> reg %s\n", nr, reg->name));
1881 /* create a new initial memory proj */
1882 assert(is_Proj(old_mem));
1883 arch_set_out_register_req(env->start, 0, arch_no_register_req);
1884 new_mem_proj = new_r_Proj(env->start, mode_M, 0);
1886 set_irg_initial_mem(irg, mem);
1888 /* Generate the Prologue */
1889 fp_reg = call->cb->prologue(env->cb, &mem, env->regs, &stack_layout->initial_bias);
1891 /* do the stack allocation BEFORE the barrier, or spill code
1892 might be added before it */
1893 env->init_sp = be_abi_reg_map_get(env->regs, sp);
1894 env->init_sp = be_new_IncSP(sp, start_bl, env->init_sp, BE_STACK_FRAME_SIZE_EXPAND, 0);
1895 be_abi_reg_map_set(env->regs, sp, env->init_sp);
1897 create_barrier(start_bl, &mem, env->regs, 0);
1899 env->init_sp = be_abi_reg_map_get(env->regs, sp);
1900 arch_set_irn_register(env->init_sp, sp);
1902 frame_pointer = be_abi_reg_map_get(env->regs, fp_reg);
1903 set_irg_frame(irg, frame_pointer);
1904 rbitset_clear(birg->allocatable_regs, fp_reg->global_index);
1906 /* rewire old mem users to new mem */
1907 exchange(old_mem, mem);
1909 /* keep the mem (for functions with an endless loop = no return) */
1912 set_irg_initial_mem(irg, mem);
1914 /* Now, introduce stack param nodes for all parameters passed on the stack */
1915 for (i = 0; i < n_params; ++i) {
1916 ir_node *arg_proj = args[i];
1917 ir_node *repl = NULL;
1919 if (arg_proj != NULL) {
1920 be_abi_call_arg_t *arg;
1921 ir_type *param_type;
1922 int nr = get_Proj_proj(arg_proj);
1925 nr = MIN(nr, n_params);
1926 arg = get_call_arg(call, 0, nr, 1);
1927 param_type = get_method_param_type(method_type, nr);
1930 repl = pmap_get(env->regs, (void *) arg->reg);
1931 } else if (arg->on_stack) {
1932 ir_node *addr = be_new_FrameAddr(sp->reg_class, start_bl, frame_pointer, arg->stack_ent);
1934 /* For atomic parameters which are actually used, we create a Load node. */
1935 if (is_atomic_type(param_type) && get_irn_n_edges(args[i]) > 0) {
1936 ir_mode *mode = get_type_mode(param_type);
1937 ir_mode *load_mode = arg->load_mode;
1939 ir_node *load = new_r_Load(start_bl, new_r_NoMem(irg), addr, load_mode, cons_floats);
1940 repl = new_r_Proj(load, load_mode, pn_Load_res);
1942 if (mode != load_mode) {
1943 repl = new_r_Conv(start_bl, repl, mode);
1946 /* The stack parameter is not primitive (it is a struct or array),
1947 * we thus will create a node representing the parameter's address
1953 assert(repl != NULL);
1955 /* Beware: the mode of the register parameters is always the mode of the register class
1956 which may be wrong. Add Conv's then. */
1957 mode = get_irn_mode(args[i]);
1958 if (mode != get_irn_mode(repl)) {
1959 repl = new_r_Conv(get_nodes_block(repl), repl, mode);
1961 exchange(args[i], repl);
1965 /* the arg proj is not needed anymore now and should be only used by the anchor */
1966 assert(get_irn_n_edges(arg_tuple) == 1);
1967 kill_node(arg_tuple);
1968 set_irg_args(irg, new_r_Bad(irg));
1970 /* All Return nodes hang on the End node, so look for them there. */
1971 end = get_irg_end_block(irg);
1972 for (i = 0, n = get_Block_n_cfgpreds(end); i < n; ++i) {
1973 ir_node *irn = get_Block_cfgpred(end, i);
1975 if (is_Return(irn)) {
1976 ir_node *blk = get_nodes_block(irn);
1977 ir_node *mem = get_Return_mem(irn);
1978 ir_node *ret = create_be_return(env, irn, blk, mem, get_Return_n_ress(irn));
1983 /* if we have endless loops here, n might be <= 0. Do NOT create a be_Return then,
1984 the code is dead and will never be executed. */
1987 /** Fix the state inputs of calls that still hang on unknowns */
1988 static void fix_call_state_inputs(ir_graph *irg)
1990 be_abi_irg_t *env = be_get_irg_abi(irg);
1991 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1993 arch_register_t **stateregs = NEW_ARR_F(arch_register_t*, 0);
1995 /* Collect caller save registers */
1996 n = arch_env->n_register_classes;
1997 for (i = 0; i < n; ++i) {
1999 const arch_register_class_t *cls = &arch_env->register_classes[i];
2000 for (j = 0; j < cls->n_regs; ++j) {
2001 const arch_register_t *reg = arch_register_for_index(cls, j);
2002 if (arch_register_type_is(reg, state)) {
2003 ARR_APP1(arch_register_t*, stateregs, (arch_register_t *)reg);
2008 n = ARR_LEN(env->calls);
2009 n_states = ARR_LEN(stateregs);
2010 for (i = 0; i < n; ++i) {
2012 ir_node *call = env->calls[i];
2014 arity = get_irn_arity(call);
2016 /* the state reg inputs are the last n inputs of the calls */
2017 for (s = 0; s < n_states; ++s) {
2018 int inp = arity - n_states + s;
2019 const arch_register_t *reg = stateregs[s];
2020 ir_node *regnode = be_abi_reg_map_get(env->regs, reg);
2022 set_irn_n(call, inp, regnode);
2026 DEL_ARR_F(stateregs);
2030 * Create a trampoline entity for the given method.
2032 static ir_entity *create_trampoline(be_main_env_t *be, ir_entity *method)
2034 ir_type *type = get_entity_type(method);
2035 ident *old_id = get_entity_ld_ident(method);
2036 ident *id = id_mangle3("", old_id, "$stub");
2037 ir_type *parent = be->pic_trampolines_type;
2038 ir_entity *ent = new_entity(parent, old_id, type);
2039 set_entity_ld_ident(ent, id);
2040 set_entity_visibility(ent, ir_visibility_private);
2046 * Returns the trampoline entity for the given method.
2048 static ir_entity *get_trampoline(be_main_env_t *env, ir_entity *method)
2050 ir_entity *result = pmap_get(env->ent_trampoline_map, method);
2051 if (result == NULL) {
2052 result = create_trampoline(env, method);
2053 pmap_insert(env->ent_trampoline_map, method, result);
2059 static ir_entity *create_pic_symbol(be_main_env_t *be, ir_entity *entity)
2061 ident *old_id = get_entity_ld_ident(entity);
2062 ident *id = id_mangle3("", old_id, "$non_lazy_ptr");
2063 ir_type *e_type = get_entity_type(entity);
2064 ir_type *type = new_type_pointer(e_type);
2065 ir_type *parent = be->pic_symbols_type;
2066 ir_entity *ent = new_entity(parent, old_id, type);
2067 set_entity_ld_ident(ent, id);
2068 set_entity_visibility(ent, ir_visibility_private);
2073 static ir_entity *get_pic_symbol(be_main_env_t *env, ir_entity *entity)
2075 ir_entity *result = pmap_get(env->ent_pic_symbol_map, entity);
2076 if (result == NULL) {
2077 result = create_pic_symbol(env, entity);
2078 pmap_insert(env->ent_pic_symbol_map, entity, result);
2087 * Returns non-zero if a given entity can be accessed using a relative address.
2089 static int can_address_relative(ir_entity *entity)
2091 return get_entity_visibility(entity) != ir_visibility_external
2092 && !(get_entity_linkage(entity) & IR_LINKAGE_MERGE);
2095 static ir_node *get_pic_base(ir_graph *irg)
2097 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
2098 if (arch_env->impl->get_pic_base == NULL)
2100 return arch_env->impl->get_pic_base(irg);
2103 /** patches SymConsts to work in position independent code */
2104 static void fix_pic_symconsts(ir_node *node, void *data)
2106 ir_graph *irg = get_irn_irg(node);
2107 be_main_env_t *be = be_get_irg_main_env(irg);
2117 arity = get_irn_arity(node);
2118 for (i = 0; i < arity; ++i) {
2120 ir_node *pred = get_irn_n(node, i);
2122 ir_entity *pic_symbol;
2123 ir_node *pic_symconst;
2125 if (!is_SymConst(pred))
2128 entity = get_SymConst_entity(pred);
2129 block = get_nodes_block(pred);
2131 /* calls can jump to relative addresses, so we can directly jump to
2132 the (relatively) known call address or the trampoline */
2133 if (i == 1 && is_Call(node)) {
2134 ir_entity *trampoline;
2135 ir_node *trampoline_const;
2137 if (can_address_relative(entity))
2140 dbgi = get_irn_dbg_info(pred);
2141 trampoline = get_trampoline(be, entity);
2142 trampoline_const = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
2144 set_irn_n(node, i, trampoline_const);
2148 /* everything else is accessed relative to EIP */
2149 mode = get_irn_mode(pred);
2150 pic_base = get_pic_base(irg);
2152 /* all ok now for locally constructed stuff */
2153 if (can_address_relative(entity)) {
2154 ir_node *add = new_r_Add(block, pic_base, pred, mode);
2156 /* make sure the walker doesn't visit this add again */
2157 mark_irn_visited(add);
2158 set_irn_n(node, i, add);
2162 /* get entry from pic symbol segment */
2163 dbgi = get_irn_dbg_info(pred);
2164 pic_symbol = get_pic_symbol(be, entity);
2165 pic_symconst = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
2167 add = new_r_Add(block, pic_base, pic_symconst, mode);
2168 mark_irn_visited(add);
2170 /* we need an extra indirection for global data outside our current
2171 module. The loads are always safe and can therefore float
2172 and need no memory input */
2173 load = new_r_Load(block, new_r_NoMem(irg), add, mode, cons_floats);
2174 load_res = new_r_Proj(load, mode, pn_Load_res);
2176 set_irn_n(node, i, load_res);
2180 be_abi_irg_t *be_abi_introduce(ir_graph *irg)
2182 be_abi_irg_t *env = XMALLOCZ(be_abi_irg_t);
2183 ir_node *old_frame = get_irg_frame(irg);
2184 be_options_t *options = be_get_irg_options(irg);
2185 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
2186 ir_entity *entity = get_irg_entity(irg);
2187 ir_type *method_type = get_entity_type(entity);
2188 be_irg_t *birg = be_birg_from_irg(irg);
2189 struct obstack *obst = &birg->obst;
2195 /* determine allocatable registers */
2196 assert(birg->allocatable_regs == NULL);
2197 birg->allocatable_regs = rbitset_obstack_alloc(obst, arch_env->n_registers);
2198 for (r = 0; r < arch_env->n_registers; ++r) {
2199 const arch_register_t *reg = &arch_env->registers[r];
2200 if ( !(reg->type & arch_register_type_ignore)) {
2201 rbitset_set(birg->allocatable_regs, r);
2205 /* break here if backend provides a custom API.
2206 * Note: we shouldn't have to setup any be_abi_irg_t* stuff at all,
2207 * but need more cleanup to make this work
2209 be_set_irg_abi(irg, env);
2211 be_omit_fp = options->omit_fp;
2213 env->dce_survivor = new_survive_dce();
2214 env->keep_map = pmap_create();
2215 env->call = be_abi_call_new(arch_env->sp->reg_class);
2216 arch_env_get_call_abi(arch_env, method_type, env->call);
2218 env->init_sp = dummy = new_r_Dummy(irg, arch_env->sp->reg_class->mode);
2219 env->calls = NEW_ARR_F(ir_node*, 0);
2222 irg_walk_graph(irg, fix_pic_symconsts, NULL, env);
2225 /* Lower all call nodes in the IRG. */
2229 Beware: init backend abi call object after processing calls,
2230 otherwise some information might be not yet available.
2232 env->cb = env->call->cb->init(env->call, irg);
2234 /* Process the IRG */
2237 /* fix call inputs for state registers */
2238 fix_call_state_inputs(irg);
2240 /* We don't need the keep map anymore. */
2241 pmap_destroy(env->keep_map);
2242 env->keep_map = NULL;
2244 /* calls array is not needed anymore */
2245 DEL_ARR_F(env->calls);
2248 /* reroute the stack origin of the calls to the true stack origin. */
2249 exchange(dummy, env->init_sp);
2250 exchange(old_frame, get_irg_frame(irg));
2252 /* Make some important node pointers survive the dead node elimination. */
2253 survive_dce_register_irn(env->dce_survivor, &env->init_sp);
2254 foreach_pmap(env->regs, ent) {
2255 survive_dce_register_irn(env->dce_survivor, (ir_node **) &ent->value);
2258 env->call->cb->done(env->cb);
2263 void be_abi_free(ir_graph *irg)
2265 be_abi_irg_t *env = be_get_irg_abi(irg);
2267 if (env->call != NULL)
2268 be_abi_call_free(env->call);
2269 if (env->dce_survivor != NULL)
2270 free_survive_dce(env->dce_survivor);
2271 if (env->regs != NULL)
2272 pmap_destroy(env->regs);
2275 be_set_irg_abi(irg, NULL);
2278 void be_put_allocatable_regs(const ir_graph *irg,
2279 const arch_register_class_t *cls, bitset_t *bs)
2281 be_irg_t *birg = be_birg_from_irg(irg);
2282 unsigned *allocatable_regs = birg->allocatable_regs;
2285 assert(bitset_size(bs) == cls->n_regs);
2286 bitset_clear_all(bs);
2287 for (i = 0; i < cls->n_regs; ++i) {
2288 const arch_register_t *reg = &cls->regs[i];
2289 if (rbitset_is_set(allocatable_regs, reg->global_index))
2294 unsigned be_get_n_allocatable_regs(const ir_graph *irg,
2295 const arch_register_class_t *cls)
2297 bitset_t *bs = bitset_alloca(cls->n_regs);
2298 be_put_allocatable_regs(irg, cls, bs);
2299 return bitset_popcount(bs);
2302 void be_set_allocatable_regs(const ir_graph *irg,
2303 const arch_register_class_t *cls,
2304 unsigned *raw_bitset)
2306 be_irg_t *birg = be_birg_from_irg(irg);
2307 unsigned *allocatable_regs = birg->allocatable_regs;
2310 rbitset_clear_all(raw_bitset, cls->n_regs);
2311 for (i = 0; i < cls->n_regs; ++i) {
2312 const arch_register_t *reg = &cls->regs[i];
2313 if (rbitset_is_set(allocatable_regs, reg->global_index))
2314 rbitset_set(raw_bitset, i);
2318 ir_node *be_abi_get_callee_save_irn(be_abi_irg_t *abi, const arch_register_t *reg)
2320 assert(arch_register_type_is(reg, callee_save));
2321 assert(pmap_contains(abi->regs, (void *) reg));
2322 return pmap_get(abi->regs, (void *) reg);
2325 ir_node *be_abi_get_ignore_irn(be_abi_irg_t *abi, const arch_register_t *reg)
2327 assert(arch_register_type_is(reg, ignore));
2328 assert(pmap_contains(abi->regs, (void *) reg));
2329 return pmap_get(abi->regs, (void *) reg);
2332 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_abi);
2333 void be_init_abi(void)
2335 FIRM_DBG_REGISTER(dbg, "firm.be.abi");