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;
365 int throws_exception;
370 /* Let the isa fill out the abi description for that call node. */
371 arch_env_get_call_abi(arch_env, call_tp, call);
373 /* Insert code to put the stack arguments on the stack. */
374 assert(get_Call_n_params(irn) == n_params);
375 stack_param_idx = ALLOCAN(int, n_params);
376 for (p = 0; p < n_params; ++p) {
377 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
380 int arg_size = get_type_size_bytes(get_method_param_type(call_tp, p));
382 stack_size += round_up2(arg->space_before, arg->alignment);
383 stack_size += round_up2(arg_size, arg->alignment);
384 stack_size += round_up2(arg->space_after, arg->alignment);
386 stack_param_idx[n_stack_params++] = p;
390 /* Collect all arguments which are passed in registers. */
391 reg_param_idxs = ALLOCAN(int, n_params);
392 for (p = 0; p < n_params; ++p) {
393 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
394 if (arg && arg->in_reg) {
395 reg_param_idxs[n_reg_params++] = p;
400 * If the stack is decreasing and we do not want to store sequentially,
401 * or someone else allocated the call frame
402 * we allocate as much space on the stack all parameters need, by
403 * moving the stack pointer along the stack's direction.
405 * Note: we also have to do this for stack_size == 0, because we may have
406 * to adjust stack alignment for the call.
408 if (!do_seq && !no_alloc) {
409 curr_sp = be_new_IncSP(sp, bl, curr_sp, stack_size, 1);
412 dbgi = get_irn_dbg_info(irn);
413 /* If there are some parameters which shall be passed on the stack. */
414 if (n_stack_params > 0) {
416 ir_node **in = ALLOCAN(ir_node*, n_stack_params+1);
420 * Reverse list of stack parameters if call arguments are from left to right.
421 * We must them reverse again if they are pushed (not stored) and the stack
422 * direction is downwards.
424 if (call->flags.bits.left_to_right ^ do_seq) {
425 for (i = 0; i < n_stack_params >> 1; ++i) {
426 int other = n_stack_params - i - 1;
427 int tmp = stack_param_idx[i];
428 stack_param_idx[i] = stack_param_idx[other];
429 stack_param_idx[other] = tmp;
433 curr_mem = get_Call_mem(irn);
435 in[n_in++] = curr_mem;
438 for (i = 0; i < n_stack_params; ++i) {
439 int p = stack_param_idx[i];
440 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
441 ir_node *param = get_Call_param(irn, p);
442 ir_node *addr = curr_sp;
444 ir_type *param_type = get_method_param_type(call_tp, p);
445 int param_size = get_type_size_bytes(param_type) + arg->space_after;
448 * If we wanted to build the arguments sequentially,
449 * the stack pointer for the next must be incremented,
450 * and the memory value propagated.
454 addr = curr_sp = be_new_IncSP(sp, bl, curr_sp,
455 param_size + arg->space_before, 0);
456 add_irn_dep(curr_sp, curr_mem);
458 curr_ofs += arg->space_before;
459 curr_ofs = round_up2(curr_ofs, arg->alignment);
461 /* Make the expression to compute the argument's offset. */
463 ir_mode *constmode = mach_mode;
464 if (mode_is_reference(mach_mode)) {
467 addr = new_r_Const_long(irg, constmode, curr_ofs);
468 addr = new_r_Add(bl, curr_sp, addr, mach_mode);
472 /* Insert a store for primitive arguments. */
473 if (is_atomic_type(param_type)) {
474 ir_node *nomem = get_irg_no_mem(irg);
475 ir_node *mem_input = do_seq ? curr_mem : nomem;
476 ir_node *store = new_rd_Store(dbgi, bl, mem_input, addr, param, cons_none);
477 mem = new_r_Proj(store, mode_M, pn_Store_M);
479 /* Make a mem copy for compound arguments. */
482 assert(mode_is_reference(get_irn_mode(param)));
483 copy = new_rd_CopyB(dbgi, bl, curr_mem, addr, param, param_type);
484 mem = new_r_Proj(copy, mode_M, pn_CopyB_M);
487 curr_ofs += param_size;
495 /* We need the sync only, if we didn't build the stores sequentially. */
497 if (n_stack_params >= 1) {
498 curr_mem = new_r_Sync(bl, n_in, in);
500 curr_mem = get_Call_mem(irn);
505 /* check for the return_twice property */
506 destroy_all_regs = 0;
507 if (is_SymConst_addr_ent(call_ptr)) {
508 ir_entity *ent = get_SymConst_entity(call_ptr);
510 if (get_entity_additional_properties(ent) & mtp_property_returns_twice)
511 destroy_all_regs = 1;
513 ir_type *call_tp = get_Call_type(irn);
515 if (get_method_additional_properties(call_tp) & mtp_property_returns_twice)
516 destroy_all_regs = 1;
519 /* Put caller save into the destroyed set and state registers in the states
521 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
523 const arch_register_class_t *cls = &arch_env->register_classes[i];
524 for (j = 0; j < cls->n_regs; ++j) {
525 const arch_register_t *reg = arch_register_for_index(cls, j);
527 /* even if destroyed all is specified, neither SP nor FP are
528 * destroyed (else bad things will happen) */
529 if (reg == arch_env->sp || reg == arch_env->bp)
532 if (reg->type & arch_register_type_state) {
533 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
534 ARR_APP1(const arch_register_t*, states, reg);
535 /* we're already in the destroyed set so no need for further
539 if (destroy_all_regs || (reg->type & arch_register_type_caller_save)) {
540 if (!(reg->type & arch_register_type_ignore)) {
541 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
547 /* search the largest result proj number */
548 res_projs = ALLOCANZ(ir_node*, n_res);
550 foreach_out_edge(irn, edge) {
551 const ir_edge_t *res_edge;
552 ir_node *irn = get_edge_src_irn(edge);
554 if (!is_Proj(irn) || get_Proj_proj(irn) != pn_Call_T_result)
557 foreach_out_edge(irn, res_edge) {
559 ir_node *res = get_edge_src_irn(res_edge);
561 assert(is_Proj(res));
563 proj = get_Proj_proj(res);
564 assert(proj < n_res);
565 assert(res_projs[proj] == NULL);
566 res_projs[proj] = res;
572 /** TODO: this is not correct for cases where return values are passed
573 * on the stack, but no known ABI does this currently...
575 n_reg_results = n_res;
578 in = ALLOCAN(ir_node*, n_reg_params + ARR_LEN(states));
580 /* make the back end call node and set its register requirements. */
581 for (i = 0; i < n_reg_params; ++i) {
582 in[n_ins++] = get_Call_param(irn, reg_param_idxs[i]);
585 /* add state registers ins */
586 for (s = 0; s < ARR_LEN(states); ++s) {
587 const arch_register_t *reg = states[s];
588 const arch_register_class_t *cls = arch_register_get_class(reg);
589 ir_node *regnode = new_r_Unknown(irg, arch_register_class_mode(cls));
590 in[n_ins++] = regnode;
592 assert(n_ins == (int) (n_reg_params + ARR_LEN(states)));
594 /* ins collected, build the call */
595 throws_exception = ir_throws_exception(irn);
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 ir_set_throws_exception(low_call, throws_exception);
609 be_Call_set_pop(low_call, call->pop);
611 /* put the call into the list of all calls for later processing */
612 ARR_APP1(ir_node *, env->calls, low_call);
614 /* create new stack pointer */
615 curr_sp = new_r_Proj(low_call, get_irn_mode(curr_sp), pn_be_Call_sp);
616 be_set_constr_single_reg_out(low_call, pn_be_Call_sp, sp,
617 arch_register_req_type_ignore | arch_register_req_type_produces_sp);
618 arch_set_irn_register(curr_sp, sp);
620 /* now handle results */
621 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);
624 long pn = i + pn_be_Call_first_res;
626 /* returns values on stack not supported yet */
630 shift the proj number to the right, since we will drop the
631 unspeakable Proj_T from the Call. Therefore, all real argument
632 Proj numbers must be increased by pn_be_Call_first_res
634 pn = i + pn_be_Call_first_res;
637 ir_type *res_type = get_method_res_type(call_tp, i);
638 ir_mode *mode = get_type_mode(res_type);
639 proj = new_r_Proj(low_call, mode, pn);
642 set_Proj_pred(proj, low_call);
643 set_Proj_proj(proj, pn);
647 /* remove register from destroyed regs */
649 size_t n = ARR_LEN(destroyed_regs);
650 for (j = 0; j < n; ++j) {
651 if (destroyed_regs[j] == arg->reg) {
652 destroyed_regs[j] = destroyed_regs[n-1];
653 ARR_SHRINKLEN(destroyed_regs,n-1);
661 Set the register class of the call address to
662 the backend provided class (default: stack pointer class)
664 be_node_set_reg_class_in(low_call, n_be_Call_ptr, call->cls_addr);
666 DBG((dbg, LEVEL_3, "\tcreated backend call %+F\n", low_call));
668 /* Set the register classes and constraints of the Call parameters. */
669 for (i = 0; i < n_reg_params; ++i) {
670 int index = reg_param_idxs[i];
671 be_abi_call_arg_t *arg = get_call_arg(call, 0, index, 0);
672 assert(arg->reg != NULL);
674 be_set_constr_single_reg_in(low_call, n_be_Call_first_arg + i,
675 arg->reg, arch_register_req_type_none);
678 /* Set the register constraints of the results. */
679 for (i = 0; i < n_res; ++i) {
680 ir_node *proj = res_projs[i];
681 const be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
682 int pn = get_Proj_proj(proj);
685 be_set_constr_single_reg_out(low_call, pn, arg->reg,
686 arch_register_req_type_none);
687 arch_set_irn_register(proj, arg->reg);
689 exchange(irn, low_call);
691 /* kill the ProjT node */
692 if (res_proj != NULL) {
696 /* Make additional projs for the caller save registers
697 and the Keep node which keeps them alive. */
703 int curr_res_proj = pn_be_Call_first_res + n_reg_results;
706 n_ins = ARR_LEN(destroyed_regs) + n_reg_results + 1;
707 in = ALLOCAN(ir_node *, n_ins);
709 /* also keep the stack pointer */
710 set_irn_link(curr_sp, (void*) sp);
713 for (d = 0; d < ARR_LEN(destroyed_regs); ++d) {
714 const arch_register_t *reg = destroyed_regs[d];
715 ir_node *proj = new_r_Proj(low_call, reg->reg_class->mode, curr_res_proj);
717 /* memorize the register in the link field. we need afterwards to set the register class of the keep correctly. */
718 be_set_constr_single_reg_out(low_call, curr_res_proj, reg,
719 arch_register_req_type_none);
720 arch_set_irn_register(proj, reg);
722 set_irn_link(proj, (void*) reg);
727 for (i = 0; i < n_reg_results; ++i) {
728 ir_node *proj = res_projs[i];
729 const arch_register_t *reg = arch_get_irn_register(proj);
730 set_irn_link(proj, (void*) reg);
735 /* create the Keep for the caller save registers */
736 keep = be_new_Keep(bl, n, in);
737 for (i = 0; i < n; ++i) {
738 const arch_register_t *reg = (const arch_register_t*)get_irn_link(in[i]);
739 be_node_set_reg_class_in(keep, i, reg->reg_class);
743 /* Clean up the stack. */
744 assert(stack_size >= call->pop);
745 stack_size -= call->pop;
747 if (stack_size > 0) {
748 ir_node *mem_proj = NULL;
750 foreach_out_edge(low_call, edge) {
751 ir_node *irn = get_edge_src_irn(edge);
752 if (is_Proj(irn) && get_Proj_proj(irn) == pn_Call_M) {
759 mem_proj = new_r_Proj(low_call, mode_M, pn_be_Call_M);
760 keep_alive(mem_proj);
763 /* Clean up the stack frame or revert alignment fixes if we allocated it */
765 curr_sp = be_new_IncSP(sp, bl, curr_sp, -stack_size, 0);
768 be_abi_call_free(call);
771 DEL_ARR_F(destroyed_regs);
777 * Adjust the size of a node representing a stack alloc or free for the minimum stack alignment.
779 * @param alignment the minimum stack alignment
780 * @param size the node containing the non-aligned size
781 * @param block the block where new nodes are allocated on
782 * @param dbg debug info for new nodes
784 * @return a node representing the aligned size
786 static ir_node *adjust_alloc_size(unsigned stack_alignment, ir_node *size,
787 ir_node *block, dbg_info *dbg)
789 if (stack_alignment > 1) {
795 assert(is_po2(stack_alignment));
797 mode = get_irn_mode(size);
798 tv = new_tarval_from_long(stack_alignment-1, mode);
799 irg = get_Block_irg(block);
800 mask = new_r_Const(irg, tv);
801 size = new_rd_Add(dbg, block, size, mask, mode);
803 tv = new_tarval_from_long(-(long)stack_alignment, mode);
804 mask = new_r_Const(irg, tv);
805 size = new_rd_And(dbg, block, size, mask, mode);
811 * The alloca is transformed into a back end alloca node and connected to the stack nodes.
813 static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp)
815 ir_node *block = get_nodes_block(alloc);
816 ir_graph *irg = get_Block_irg(block);
817 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
818 ir_node *alloc_mem = NULL;
819 ir_node *alloc_res = NULL;
820 ir_type *type = get_Alloc_type(alloc);
823 const ir_edge_t *edge;
828 unsigned stack_alignment;
830 /* all non-stack Alloc nodes should already be lowered before the backend */
831 assert(get_Alloc_where(alloc) == stack_alloc);
833 foreach_out_edge(alloc, edge) {
834 ir_node *irn = get_edge_src_irn(edge);
836 assert(is_Proj(irn));
837 switch (get_Proj_proj(irn)) {
849 /* Beware: currently Alloc nodes without a result might happen,
850 only escape analysis kills them and this phase runs only for object
851 oriented source. We kill the Alloc here. */
852 if (alloc_res == NULL && alloc_mem) {
853 exchange(alloc_mem, get_Alloc_mem(alloc));
857 dbg = get_irn_dbg_info(alloc);
858 count = get_Alloc_count(alloc);
860 /* we might need to multiply the count with the element size */
861 if (type != firm_unknown_type && get_type_size_bytes(type) != 1) {
862 ir_mode *mode = get_irn_mode(count);
863 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type),
865 ir_node *cnst = new_rd_Const(dbg, irg, tv);
866 size = new_rd_Mul(dbg, block, count, cnst, mode);
871 /* The stack pointer will be modified in an unknown manner.
872 We cannot omit it. */
873 env->call->flags.bits.try_omit_fp = 0;
875 stack_alignment = 1 << arch_env->stack_alignment;
876 size = adjust_alloc_size(stack_alignment, size, block, dbg);
877 new_alloc = be_new_AddSP(arch_env->sp, block, curr_sp, size);
878 set_irn_dbg_info(new_alloc, dbg);
880 if (alloc_mem != NULL) {
884 addsp_mem = new_r_Proj(new_alloc, mode_M, pn_be_AddSP_M);
886 /* We need to sync the output mem of the AddSP with the input mem
887 edge into the alloc node. */
888 ins[0] = get_Alloc_mem(alloc);
890 sync = new_r_Sync(block, 2, ins);
892 exchange(alloc_mem, sync);
895 exchange(alloc, new_alloc);
897 /* fix projnum of alloca res */
898 set_Proj_proj(alloc_res, pn_be_AddSP_res);
900 curr_sp = new_r_Proj(new_alloc, get_irn_mode(curr_sp), pn_be_AddSP_sp);
907 * The Free is transformed into a back end free node and connected to the stack nodes.
909 static ir_node *adjust_free(be_abi_irg_t *env, ir_node *free, ir_node *curr_sp)
911 ir_node *block = get_nodes_block(free);
912 ir_graph *irg = get_irn_irg(free);
913 ir_type *type = get_Free_type(free);
914 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
915 ir_mode *sp_mode = arch_env->sp->reg_class->mode;
916 dbg_info *dbg = get_irn_dbg_info(free);
917 ir_node *subsp, *mem, *res, *size, *sync;
919 unsigned stack_alignment;
921 /* all non-stack-alloc Free nodes should already be lowered before the
923 assert(get_Free_where(free) == stack_alloc);
925 /* we might need to multiply the size with the element size */
926 if (type != firm_unknown_type && get_type_size_bytes(type) != 1) {
927 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type), mode_Iu);
928 ir_node *cnst = new_rd_Const(dbg, irg, tv);
929 ir_node *mul = new_rd_Mul(dbg, block, get_Free_size(free),
933 size = get_Free_size(free);
936 stack_alignment = 1 << arch_env->stack_alignment;
937 size = adjust_alloc_size(stack_alignment, size, block, dbg);
939 /* The stack pointer will be modified in an unknown manner.
940 We cannot omit it. */
941 env->call->flags.bits.try_omit_fp = 0;
942 subsp = be_new_SubSP(arch_env->sp, block, curr_sp, size);
943 set_irn_dbg_info(subsp, dbg);
945 mem = new_r_Proj(subsp, mode_M, pn_be_SubSP_M);
946 res = new_r_Proj(subsp, sp_mode, pn_be_SubSP_sp);
948 /* we need to sync the memory */
949 in[0] = get_Free_mem(free);
951 sync = new_r_Sync(block, 2, in);
953 /* and make the AddSP dependent on the former memory */
954 add_irn_dep(subsp, get_Free_mem(free));
957 exchange(free, sync);
964 * Check if a node is somehow data dependent on another one.
965 * both nodes must be in the same basic block.
966 * @param n1 The first node.
967 * @param n2 The second node.
968 * @return 1, if n1 is data dependent (transitively) on n2, 0 if not.
970 static int dependent_on(ir_node *n1, ir_node *n2)
972 assert(get_nodes_block(n1) == get_nodes_block(n2));
974 return heights_reachable_in_block(ir_heights, n1, n2);
977 static int cmp_call_dependency(const void *c1, const void *c2)
979 ir_node *n1 = *(ir_node **) c1;
980 ir_node *n2 = *(ir_node **) c2;
984 Classical qsort() comparison function behavior:
985 0 if both elements are equal
986 1 if second is "smaller" that first
987 -1 if first is "smaller" that second
989 if (dependent_on(n1, n2))
992 if (dependent_on(n2, n1))
995 /* The nodes have no depth order, but we need a total order because qsort()
998 * Additionally, we need to respect transitive dependencies. Consider a
999 * Call a depending on Call b and an independent Call c.
1000 * We MUST NOT order c > a and b > c. */
1001 h1 = get_irn_height(ir_heights, n1);
1002 h2 = get_irn_height(ir_heights, n2);
1003 if (h1 < h2) return -1;
1004 if (h1 > h2) return 1;
1005 /* Same height, so use a random (but stable) order */
1006 return get_irn_idx(n1) - get_irn_idx(n2);
1010 * Walker: links all Call/Alloc/Free nodes to the Block they are contained.
1011 * Clears the irg_is_leaf flag if a Call is detected.
1013 static void link_ops_in_block_walker(ir_node *irn, void *data)
1015 be_abi_irg_t *env = (be_abi_irg_t*)data;
1016 unsigned code = get_irn_opcode(irn);
1018 if (code == iro_Call ||
1019 (code == iro_Alloc && get_Alloc_where(irn) == stack_alloc) ||
1020 (code == iro_Free && get_Free_where(irn) == stack_alloc)) {
1021 ir_node *bl = get_nodes_block(irn);
1022 void *save = get_irn_link(bl);
1024 if (code == iro_Call)
1025 env->call->flags.bits.irg_is_leaf = 0;
1027 set_irn_link(irn, save);
1028 set_irn_link(bl, irn);
1031 if (code == iro_Builtin && get_Builtin_kind(irn) == ir_bk_return_address) {
1032 ir_node *param = get_Builtin_param(irn, 0);
1033 ir_tarval *tv = get_Const_tarval(param);
1034 unsigned long value = get_tarval_long(tv);
1035 /* use ebp, so the climbframe algo works... */
1037 env->call->flags.bits.try_omit_fp = 0;
1044 * Process all Call/Alloc/Free nodes inside a basic block.
1045 * Note that the link field of the block must contain a linked list of all
1046 * nodes inside the Block. We first order this list according to data dependency
1047 * and that connect the nodes together.
1049 static void process_ops_in_block(ir_node *bl, void *data)
1051 be_abi_irg_t *env = (be_abi_irg_t*)data;
1052 ir_node *curr_sp = env->init_sp;
1059 for (irn = (ir_node*)get_irn_link(bl); irn != NULL;
1060 irn = (ir_node*)get_irn_link(irn)) {
1064 nodes = ALLOCAN(ir_node*, n_nodes);
1065 for (irn = (ir_node*)get_irn_link(bl), n = 0; irn != NULL;
1066 irn = (ir_node*)get_irn_link(irn), ++n) {
1070 /* If there were call nodes in the block. */
1075 /* order the call nodes according to data dependency */
1076 qsort(nodes, n_nodes, sizeof(nodes[0]), cmp_call_dependency);
1078 for (i = n_nodes - 1; i >= 0; --i) {
1079 ir_node *irn = nodes[i];
1081 DBG((dbg, LEVEL_3, "\tprocessing call %+F\n", irn));
1082 switch (get_irn_opcode(irn)) {
1085 /* The stack pointer will be modified due to a call. */
1086 env->call->flags.bits.try_omit_fp = 0;
1088 curr_sp = adjust_call(env, irn, curr_sp);
1091 if (get_Alloc_where(irn) == stack_alloc)
1092 curr_sp = adjust_alloc(env, irn, curr_sp);
1095 if (get_Free_where(irn) == stack_alloc)
1096 curr_sp = adjust_free(env, irn, curr_sp);
1099 panic("invalid call");
1103 /* Keep the last stack state in the block by tying it to Keep node,
1104 * the proj from calls is already kept */
1105 if (curr_sp != env->init_sp &&
1106 !(is_Proj(curr_sp) && be_is_Call(get_Proj_pred(curr_sp)))) {
1108 keep = be_new_Keep(bl, 1, nodes);
1109 pmap_insert(env->keep_map, bl, keep);
1113 set_irn_link(bl, curr_sp);
1117 * Adjust all call nodes in the graph to the ABI conventions.
1119 static void process_calls(ir_graph *irg)
1121 be_abi_irg_t *abi = be_get_irg_abi(irg);
1123 abi->call->flags.bits.irg_is_leaf = 1;
1124 irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, abi);
1126 ir_heights = heights_new(irg);
1127 irg_block_walk_graph(irg, NULL, process_ops_in_block, abi);
1128 heights_free(ir_heights);
1132 * Computes the stack argument layout type.
1133 * Changes a possibly allocated value param type by moving
1134 * entities to the stack layout type.
1136 * @param env the ABI environment
1137 * @param call the current call ABI
1138 * @param method_type the method type
1139 * @param val_param_tp the value parameter type, will be destroyed
1140 * @param param_map an array mapping method arguments to the stack layout type
1142 * @return the stack argument layout type
1144 static ir_type *compute_arg_type(be_abi_irg_t *env, ir_graph *irg,
1145 be_abi_call_t *call,
1146 ir_type *method_type, ir_type *val_param_tp,
1147 ir_entity ***param_map)
1149 int dir = env->call->flags.bits.left_to_right ? 1 : -1;
1151 int n = get_method_n_params(method_type);
1152 int curr = inc > 0 ? 0 : n - 1;
1153 struct obstack *obst = be_get_be_obst(irg);
1159 ident *id = get_entity_ident(get_irg_entity(irg));
1162 *param_map = map = OALLOCN(obst, ir_entity*, n);
1163 res = new_type_struct(id_mangle_u(id, new_id_from_chars("arg_type", 8)));
1164 for (i = 0; i < n; ++i, curr += inc) {
1165 ir_type *param_type = get_method_param_type(method_type, curr);
1166 be_abi_call_arg_t *arg = get_call_arg(call, 0, curr, 1);
1169 if (arg->on_stack) {
1170 if (val_param_tp != NULL) {
1171 /* the entity was already created, create a copy in the param type */
1172 ir_entity *val_ent = get_method_value_param_ent(method_type, i);
1173 arg->stack_ent = copy_entity_own(val_ent, res);
1174 set_entity_link(val_ent, arg->stack_ent);
1175 set_entity_link(arg->stack_ent, NULL);
1177 /* create a new entity */
1178 snprintf(buf, sizeof(buf), "param_%d", i);
1179 arg->stack_ent = new_entity(res, new_id_from_str(buf), param_type);
1181 ofs += arg->space_before;
1182 ofs = round_up2(ofs, arg->alignment);
1183 set_entity_offset(arg->stack_ent, ofs);
1184 ofs += arg->space_after;
1185 ofs += get_type_size_bytes(param_type);
1186 map[i] = arg->stack_ent;
1189 set_type_size_bytes(res, ofs);
1190 set_type_state(res, layout_fixed);
1195 const arch_register_t *reg;
1199 static int cmp_regs(const void *a, const void *b)
1201 const reg_node_map_t *p = (const reg_node_map_t*)a;
1202 const reg_node_map_t *q = (const reg_node_map_t*)b;
1204 if (p->reg->reg_class == q->reg->reg_class)
1205 return p->reg->index - q->reg->index;
1207 return p->reg->reg_class < q->reg->reg_class ? -1 : +1;
1210 static void reg_map_to_arr(reg_node_map_t *res, pmap *reg_map)
1213 size_t n = pmap_count(reg_map);
1216 foreach_pmap(reg_map, ent) {
1217 res[i].reg = (const arch_register_t*)ent->key;
1218 res[i].irn = (ir_node*)ent->value;
1222 qsort(res, n, sizeof(res[0]), cmp_regs);
1226 * Creates a be_Return for a Return node.
1228 * @param @env the abi environment
1229 * @param irn the Return node or NULL if there was none
1230 * @param bl the block where the be_Retun should be placed
1231 * @param mem the current memory
1232 * @param n_res number of return results
1234 static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl,
1235 ir_node *mem, int n_res)
1237 be_abi_call_t *call = env->call;
1238 ir_graph *irg = get_Block_irg(bl);
1239 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1241 pmap *reg_map = pmap_create();
1242 ir_node *keep = (ir_node*)pmap_get(env->keep_map, bl);
1249 const arch_register_t **regs;
1253 get the valid stack node in this block.
1254 If we had a call in that block there is a Keep constructed by process_calls()
1255 which points to the last stack modification in that block. we'll use
1256 it then. Else we use the stack from the start block and let
1257 the ssa construction fix the usage.
1259 stack = be_abi_reg_map_get(env->regs, arch_env->sp);
1261 stack = get_irn_n(keep, 0);
1263 remove_End_keepalive(get_irg_end(irg), keep);
1266 /* Insert results for Return into the register map. */
1267 for (i = 0; i < n_res; ++i) {
1268 ir_node *res = get_Return_res(irn, i);
1269 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1270 assert(arg->in_reg && "return value must be passed in register");
1271 pmap_insert(reg_map, (void *) arg->reg, res);
1274 /* Add uses of the callee save registers. */
1275 foreach_pmap(env->regs, ent) {
1276 const arch_register_t *reg = (const arch_register_t*)ent->key;
1277 if (reg->type & (arch_register_type_callee_save | arch_register_type_ignore))
1278 pmap_insert(reg_map, ent->key, ent->value);
1281 be_abi_reg_map_set(reg_map, arch_env->sp, stack);
1284 Maximum size of the in array for Return nodes is
1285 return args + callee save/ignore registers + memory + stack pointer
1287 in_max = pmap_count(reg_map) + n_res + 2;
1289 in = ALLOCAN(ir_node*, in_max);
1290 regs = ALLOCAN(arch_register_t const*, in_max);
1293 in[1] = be_abi_reg_map_get(reg_map, arch_env->sp);
1295 regs[1] = arch_env->sp;
1298 /* clear SP entry, since it has already been grown. */
1299 pmap_insert(reg_map, (void *) arch_env->sp, NULL);
1300 for (i = 0; i < n_res; ++i) {
1301 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1303 in[n] = be_abi_reg_map_get(reg_map, arg->reg);
1304 regs[n++] = arg->reg;
1306 /* Clear the map entry to mark the register as processed. */
1307 be_abi_reg_map_set(reg_map, arg->reg, NULL);
1310 /* grow the rest of the stuff. */
1311 foreach_pmap(reg_map, ent) {
1313 in[n] = (ir_node*)ent->value;
1314 regs[n++] = (const arch_register_t*)ent->key;
1318 /* The in array for the new back end return is now ready. */
1320 dbgi = get_irn_dbg_info(irn);
1324 /* we have to pop the shadow parameter in in case of struct returns */
1326 ret = be_new_Return(dbgi, irg, bl, n_res, pop, n, in);
1328 /* Set the register classes of the return's parameter accordingly. */
1329 for (i = 0; i < n; ++i) {
1330 if (regs[i] == NULL)
1333 be_set_constr_single_reg_in(ret, i, regs[i], arch_register_req_type_none);
1336 /* Free the space of the Epilog's in array and the register <-> proj map. */
1337 pmap_destroy(reg_map);
1342 typedef struct ent_pos_pair ent_pos_pair;
1343 struct ent_pos_pair {
1344 ir_entity *ent; /**< a value param entity */
1345 int pos; /**< its parameter number */
1346 ent_pos_pair *next; /**< for linking */
1349 typedef struct lower_frame_sels_env_t {
1350 ent_pos_pair *value_param_list; /**< the list of all value param entities */
1351 ir_node *frame; /**< the current frame */
1352 const arch_register_class_t *sp_class; /**< register class of the stack pointer */
1353 const arch_register_class_t *link_class; /**< register class of the link pointer */
1354 ir_type *value_tp; /**< the value type if any */
1355 ir_type *frame_tp; /**< the frame type */
1356 int static_link_pos; /**< argument number of the hidden static link */
1357 } lower_frame_sels_env_t;
1360 * Return an entity from the backend for an value param entity.
1362 * @param ent an value param type entity
1363 * @param ctx context
1365 static ir_entity *get_argument_entity(ir_entity *ent, lower_frame_sels_env_t *ctx)
1367 ir_entity *argument_ent = (ir_entity*)get_entity_link(ent);
1369 if (argument_ent == NULL) {
1370 /* we have NO argument entity yet: This is bad, as we will
1371 * need one for backing store.
1374 ir_type *frame_tp = ctx->frame_tp;
1375 unsigned offset = get_type_size_bytes(frame_tp);
1376 ir_type *tp = get_entity_type(ent);
1377 unsigned align = get_type_alignment_bytes(tp);
1379 offset += align - 1;
1380 offset &= ~(align - 1);
1382 argument_ent = copy_entity_own(ent, frame_tp);
1384 /* must be automatic to set a fixed layout */
1385 set_entity_offset(argument_ent, offset);
1386 offset += get_type_size_bytes(tp);
1388 set_type_size_bytes(frame_tp, offset);
1389 set_entity_link(ent, argument_ent);
1391 return argument_ent;
1394 * Walker: Replaces Sels of frame type and
1395 * value param type entities by FrameAddress.
1396 * Links all used entities.
1398 static void lower_frame_sels_walker(ir_node *irn, void *data)
1400 lower_frame_sels_env_t *ctx = (lower_frame_sels_env_t*)data;
1403 ir_node *ptr = get_Sel_ptr(irn);
1405 if (ptr == ctx->frame) {
1406 ir_entity *ent = get_Sel_entity(irn);
1407 ir_node *bl = get_nodes_block(irn);
1410 int is_value_param = 0;
1412 if (get_entity_owner(ent) == ctx->value_tp) {
1415 /* replace by its copy from the argument type */
1416 pos = get_struct_member_index(ctx->value_tp, ent);
1417 ent = get_argument_entity(ent, ctx);
1420 nw = be_new_FrameAddr(ctx->sp_class, bl, ctx->frame, ent);
1423 /* check, if it's a param Sel and if have not seen this entity before */
1424 if (is_value_param && get_entity_link(ent) == NULL) {
1430 ARR_APP1(ent_pos_pair, ctx->value_param_list, pair);
1432 set_entity_link(ent, ctx->value_param_list);
1439 * Check if a value parameter is transmitted as a register.
1440 * This might happen if the address of an parameter is taken which is
1441 * transmitted in registers.
1443 * Note that on some architectures this case must be handled specially
1444 * because the place of the backing store is determined by their ABI.
1446 * In the default case we move the entity to the frame type and create
1447 * a backing store into the first block.
1449 static void fix_address_of_parameter_access(be_abi_irg_t *env, ir_graph *irg,
1450 ent_pos_pair *value_param_list)
1452 be_abi_call_t *call = env->call;
1453 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1454 ent_pos_pair *entry, *new_list;
1456 int i, n = ARR_LEN(value_param_list);
1459 for (i = 0; i < n; ++i) {
1460 int pos = value_param_list[i].pos;
1461 be_abi_call_arg_t *arg = get_call_arg(call, 0, pos, 1);
1464 DBG((dbg, LEVEL_2, "\targ #%d need backing store\n", pos));
1465 value_param_list[i].next = new_list;
1466 new_list = &value_param_list[i];
1469 if (new_list != NULL) {
1470 /* ok, change the graph */
1471 ir_node *start_bl = get_irg_start_block(irg);
1472 ir_node *first_bl = get_first_block_succ(start_bl);
1473 ir_node *frame, *imem, *nmem, *store, *mem, *args;
1474 optimization_state_t state;
1477 assert(first_bl && first_bl != start_bl);
1478 /* we had already removed critical edges, so the following
1479 assertion should be always true. */
1480 assert(get_Block_n_cfgpreds(first_bl) == 1);
1482 /* now create backing stores */
1483 frame = get_irg_frame(irg);
1484 imem = get_irg_initial_mem(irg);
1486 save_optimization_state(&state);
1488 nmem = new_r_Proj(get_irg_start(irg), mode_M, pn_Start_M);
1489 restore_optimization_state(&state);
1491 /* reroute all edges to the new memory source */
1492 edges_reroute(imem, nmem);
1496 args = get_irg_args(irg);
1497 for (entry = new_list; entry != NULL; entry = entry->next) {
1499 ir_type *tp = get_entity_type(entry->ent);
1500 ir_mode *mode = get_type_mode(tp);
1503 /* address for the backing store */
1504 addr = be_new_FrameAddr(arch_env->sp->reg_class, first_bl, frame, entry->ent);
1507 mem = new_r_Proj(store, mode_M, pn_Store_M);
1509 /* the backing store itself */
1510 store = new_r_Store(first_bl, mem, addr,
1511 new_r_Proj(args, mode, i), cons_none);
1513 /* the new memory Proj gets the last Proj from store */
1514 set_Proj_pred(nmem, store);
1515 set_Proj_proj(nmem, pn_Store_M);
1516 set_nodes_block(nmem, get_nodes_block(store));
1518 /* move all entities to the frame type */
1519 frame_tp = get_irg_frame_type(irg);
1520 offset = get_type_size_bytes(frame_tp);
1522 /* we will add new entities: set the layout to undefined */
1523 assert(get_type_state(frame_tp) == layout_fixed);
1524 set_type_state(frame_tp, layout_undefined);
1525 for (entry = new_list; entry != NULL; entry = entry->next) {
1526 ir_entity *ent = entry->ent;
1528 /* If the entity is still on the argument type, move it to the
1530 * This happens if the value_param type was build due to compound
1532 if (get_entity_owner(ent) != frame_tp) {
1533 ir_type *tp = get_entity_type(ent);
1534 unsigned align = get_type_alignment_bytes(tp);
1536 offset += align - 1;
1537 offset &= ~(align - 1);
1538 set_entity_owner(ent, frame_tp);
1539 /* must be automatic to set a fixed layout */
1540 set_entity_offset(ent, offset);
1541 offset += get_type_size_bytes(tp);
1544 set_type_size_bytes(frame_tp, offset);
1545 /* fix the layout again */
1546 set_type_state(frame_tp, layout_fixed);
1551 * The start block has no jump, instead it has an initial exec Proj.
1552 * The backend wants to handle all blocks the same way, so we replace
1553 * the out cfg edge with a real jump.
1555 static void fix_start_block(ir_graph *irg)
1557 ir_node *initial_X = get_irg_initial_exec(irg);
1558 ir_node *start_block = get_irg_start_block(irg);
1559 ir_node *jmp = new_r_Jmp(start_block);
1561 assert(is_Proj(initial_X));
1562 exchange(initial_X, jmp);
1563 set_irg_initial_exec(irg, new_r_Bad(irg, mode_X));
1565 /* merge start block with successor if possible */
1567 const ir_edge_t *edge;
1568 foreach_out_edge(jmp, edge) {
1569 ir_node *succ = get_edge_src_irn(edge);
1570 if (!is_Block(succ))
1573 if (get_irn_arity(succ) == 1) {
1574 exchange(succ, start_block);
1582 * Update the entity of Sels to the outer value parameters.
1584 static void update_outer_frame_sels(ir_node *irn, void *env)
1586 lower_frame_sels_env_t *ctx = (lower_frame_sels_env_t*)env;
1593 ptr = get_Sel_ptr(irn);
1594 if (! is_arg_Proj(ptr))
1596 if (get_Proj_proj(ptr) != ctx->static_link_pos)
1598 ent = get_Sel_entity(irn);
1600 if (get_entity_owner(ent) == ctx->value_tp) {
1601 /* replace by its copy from the argument type */
1602 pos = get_struct_member_index(ctx->value_tp, ent);
1603 ent = get_argument_entity(ent, ctx);
1604 set_Sel_entity(irn, ent);
1606 /* check, if we have not seen this entity before */
1607 if (get_entity_link(ent) == NULL) {
1613 ARR_APP1(ent_pos_pair, ctx->value_param_list, pair);
1615 set_entity_link(ent, ctx->value_param_list);
1621 * Fix access to outer local variables.
1623 static void fix_outer_variable_access(be_abi_irg_t *env,
1624 lower_frame_sels_env_t *ctx)
1630 for (i = get_class_n_members(ctx->frame_tp) - 1; i >= 0; --i) {
1631 ir_entity *ent = get_class_member(ctx->frame_tp, i);
1633 if (! is_method_entity(ent))
1636 irg = get_entity_irg(ent);
1641 * FIXME: find the number of the static link parameter
1642 * for now we assume 0 here
1644 ctx->static_link_pos = 0;
1646 irg_walk_graph(irg, NULL, update_outer_frame_sels, ctx);
1651 * Modify the irg itself and the frame type.
1653 static void modify_irg(ir_graph *irg)
1655 be_abi_irg_t *env = be_get_irg_abi(irg);
1656 be_abi_call_t *call = env->call;
1657 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1658 const arch_register_t *sp = arch_env->sp;
1659 ir_type *method_type = get_entity_type(get_irg_entity(irg));
1660 be_irg_t *birg = be_birg_from_irg(irg);
1661 struct obstack *obst = be_get_be_obst(irg);
1662 be_stack_layout_t *stack_layout = be_get_irg_stack_layout(irg);
1665 ir_node *new_mem_proj;
1671 unsigned frame_size;
1674 const arch_register_t *fp_reg;
1675 ir_node *frame_pointer;
1679 const ir_edge_t *edge;
1680 ir_type *arg_type, *bet_type, *tp;
1681 lower_frame_sels_env_t ctx;
1682 ir_entity **param_map;
1684 DBG((dbg, LEVEL_1, "introducing abi on %+F\n", irg));
1686 old_mem = get_irg_initial_mem(irg);
1688 irp_reserve_resources(irp, IR_RESOURCE_ENTITY_LINK);
1690 /* set the links of all frame entities to NULL, we use it
1691 to detect if an entity is already linked in the value_param_list */
1692 tp = get_method_value_param_type(method_type);
1695 /* clear the links of the clone type, let the
1696 original entities point to its clones */
1697 for (i = get_struct_n_members(tp) - 1; i >= 0; --i) {
1698 ir_entity *mem = get_struct_member(tp, i);
1699 set_entity_link(mem, NULL);
1703 arg_type = compute_arg_type(env, irg, call, method_type, tp, ¶m_map);
1705 /* Convert the Sel nodes in the irg to frame addr nodes: */
1706 ctx.value_param_list = NEW_ARR_F(ent_pos_pair, 0);
1707 ctx.frame = get_irg_frame(irg);
1708 ctx.sp_class = arch_env->sp->reg_class;
1709 ctx.link_class = arch_env->link_class;
1710 ctx.frame_tp = get_irg_frame_type(irg);
1712 /* layout the stackframe now */
1713 if (get_type_state(ctx.frame_tp) == layout_undefined) {
1714 default_layout_compound_type(ctx.frame_tp);
1717 /* we will possible add new entities to the frame: set the layout to undefined */
1718 assert(get_type_state(ctx.frame_tp) == layout_fixed);
1719 set_type_state(ctx.frame_tp, layout_undefined);
1721 irg_walk_graph(irg, lower_frame_sels_walker, NULL, &ctx);
1723 /* fix the frame type layout again */
1724 set_type_state(ctx.frame_tp, layout_fixed);
1725 /* align stackframe to 4 byte */
1726 frame_size = get_type_size_bytes(ctx.frame_tp);
1727 if (frame_size % 4 != 0) {
1728 set_type_size_bytes(ctx.frame_tp, frame_size + 4 - (frame_size % 4));
1731 env->regs = pmap_create();
1733 n_params = get_method_n_params(method_type);
1734 args = OALLOCNZ(obst, ir_node*, n_params);
1737 * for inner function we must now fix access to outer frame entities.
1739 fix_outer_variable_access(env, &ctx);
1741 /* Check if a value parameter is transmitted as a register.
1742 * This might happen if the address of an parameter is taken which is
1743 * transmitted in registers.
1745 * Note that on some architectures this case must be handled specially
1746 * because the place of the backing store is determined by their ABI.
1748 * In the default case we move the entity to the frame type and create
1749 * a backing store into the first block.
1751 fix_address_of_parameter_access(env, irg, ctx.value_param_list);
1753 DEL_ARR_F(ctx.value_param_list);
1754 irp_free_resources(irp, IR_RESOURCE_ENTITY_LINK);
1756 /* Fill the argument vector */
1757 arg_tuple = get_irg_args(irg);
1758 foreach_out_edge(arg_tuple, edge) {
1759 ir_node *irn = get_edge_src_irn(edge);
1760 if (! is_Anchor(irn)) {
1761 int nr = get_Proj_proj(irn);
1763 DBG((dbg, LEVEL_2, "\treading arg: %d -> %+F\n", nr, irn));
1767 stack_layout->sp_relative = call->flags.bits.try_omit_fp;
1768 bet_type = call->cb->get_between_type(irg);
1769 stack_frame_init(stack_layout, arg_type, bet_type,
1770 get_irg_frame_type(irg), param_map);
1772 /* Count the register params and add them to the number of Projs for the RegParams node */
1773 for (i = 0; i < n_params; ++i) {
1774 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1775 if (arg->in_reg && args[i]) {
1776 assert(arg->reg != sp && "cannot use stack pointer as parameter register");
1777 assert(i == get_Proj_proj(args[i]));
1779 /* For now, associate the register with the old Proj from Start representing that argument. */
1780 pmap_insert(env->regs, (void *) arg->reg, args[i]);
1781 DBG((dbg, LEVEL_2, "\targ #%d -> reg %s\n", i, arg->reg->name));
1785 /* Collect all callee-save registers */
1786 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
1787 const arch_register_class_t *cls = &arch_env->register_classes[i];
1788 for (j = 0; j < cls->n_regs; ++j) {
1789 const arch_register_t *reg = &cls->regs[j];
1790 if (reg->type & (arch_register_type_callee_save | arch_register_type_state)) {
1791 pmap_insert(env->regs, (void *) reg, NULL);
1796 fp_reg = call->flags.bits.try_omit_fp ? arch_env->sp : arch_env->bp;
1797 rbitset_clear(birg->allocatable_regs, fp_reg->global_index);
1799 /* handle start block here (place a jump in the block) */
1800 fix_start_block(irg);
1802 pmap_insert(env->regs, (void *) sp, NULL);
1803 pmap_insert(env->regs, (void *) arch_env->bp, NULL);
1804 start_bl = get_irg_start_block(irg);
1805 env->start = be_new_Start(NULL, start_bl, pmap_count(env->regs) + 1);
1806 set_irg_start(irg, env->start);
1809 * make proj nodes for the callee save registers.
1810 * memorize them, since Return nodes get those as inputs.
1812 * Note, that if a register corresponds to an argument, the regs map
1813 * contains the old Proj from start for that argument.
1815 rm = ALLOCAN(reg_node_map_t, pmap_count(env->regs));
1816 reg_map_to_arr(rm, env->regs);
1817 for (i = 0, n = pmap_count(env->regs); i < n; ++i) {
1818 const arch_register_t *reg = rm[i].reg;
1819 ir_mode *mode = reg->reg_class->mode;
1821 arch_register_req_type_t add_type = arch_register_req_type_none;
1825 add_type |= arch_register_req_type_produces_sp;
1826 if (!rbitset_is_set(birg->allocatable_regs, reg->global_index)) {
1827 add_type |= arch_register_req_type_ignore;
1831 proj = new_r_Proj(env->start, mode, nr + 1);
1832 pmap_insert(env->regs, (void *) reg, proj);
1833 be_set_constr_single_reg_out(env->start, nr + 1, reg, add_type);
1834 arch_set_irn_register(proj, reg);
1836 DBG((dbg, LEVEL_2, "\tregister save proj #%d -> reg %s\n", nr, reg->name));
1839 /* create a new initial memory proj */
1840 assert(is_Proj(old_mem));
1841 arch_set_out_register_req(env->start, 0, arch_no_register_req);
1842 new_mem_proj = new_r_Proj(env->start, mode_M, 0);
1844 set_irg_initial_mem(irg, mem);
1846 env->init_sp = be_abi_reg_map_get(env->regs, sp);
1848 /* set new frame_pointer */
1849 frame_pointer = be_abi_reg_map_get(env->regs, fp_reg);
1850 set_irg_frame(irg, frame_pointer);
1852 /* rewire old mem users to new mem */
1853 exchange(old_mem, mem);
1855 /* keep the mem (for functions with an endless loop = no return) */
1858 set_irg_initial_mem(irg, mem);
1860 /* Now, introduce stack param nodes for all parameters passed on the stack */
1861 for (i = 0; i < n_params; ++i) {
1862 ir_node *arg_proj = args[i];
1863 ir_node *repl = NULL;
1865 if (arg_proj != NULL) {
1866 be_abi_call_arg_t *arg;
1867 ir_type *param_type;
1868 int nr = get_Proj_proj(arg_proj);
1871 nr = MIN(nr, n_params);
1872 arg = get_call_arg(call, 0, nr, 1);
1873 param_type = get_method_param_type(method_type, nr);
1876 repl = (ir_node*)pmap_get(env->regs, arg->reg);
1877 } else if (arg->on_stack) {
1878 ir_node *addr = be_new_FrameAddr(sp->reg_class, start_bl, frame_pointer, arg->stack_ent);
1880 /* For atomic parameters which are actually used, we create a Load node. */
1881 if (is_atomic_type(param_type) && get_irn_n_edges(args[i]) > 0) {
1882 ir_mode *mode = get_type_mode(param_type);
1883 ir_mode *load_mode = arg->load_mode;
1884 ir_node *nomem = get_irg_no_mem(irg);
1886 ir_node *load = new_r_Load(start_bl, nomem, addr, load_mode, cons_floats);
1887 repl = new_r_Proj(load, load_mode, pn_Load_res);
1889 if (mode != load_mode) {
1890 repl = new_r_Conv(start_bl, repl, mode);
1893 /* The stack parameter is not primitive (it is a struct or array),
1894 * we thus will create a node representing the parameter's address
1900 assert(repl != NULL);
1902 /* Beware: the mode of the register parameters is always the mode of the register class
1903 which may be wrong. Add Conv's then. */
1904 mode = get_irn_mode(args[i]);
1905 if (mode != get_irn_mode(repl)) {
1906 repl = new_r_Conv(get_nodes_block(repl), repl, mode);
1908 exchange(args[i], repl);
1912 /* the arg proj is not needed anymore now and should be only used by the anchor */
1913 assert(get_irn_n_edges(arg_tuple) == 1);
1914 kill_node(arg_tuple);
1915 set_irg_args(irg, new_r_Bad(irg, mode_T));
1917 /* All Return nodes hang on the End node, so look for them there. */
1918 end = get_irg_end_block(irg);
1919 for (i = 0, n = get_Block_n_cfgpreds(end); i < n; ++i) {
1920 ir_node *irn = get_Block_cfgpred(end, i);
1922 if (is_Return(irn)) {
1923 ir_node *blk = get_nodes_block(irn);
1924 ir_node *mem = get_Return_mem(irn);
1925 ir_node *ret = create_be_return(env, irn, blk, mem, get_Return_n_ress(irn));
1930 /* if we have endless loops here, n might be <= 0. Do NOT create a be_Return then,
1931 the code is dead and will never be executed. */
1934 /** Fix the state inputs of calls that still hang on unknowns */
1935 static void fix_call_state_inputs(ir_graph *irg)
1937 be_abi_irg_t *env = be_get_irg_abi(irg);
1938 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1940 arch_register_t **stateregs = NEW_ARR_F(arch_register_t*, 0);
1942 /* Collect caller save registers */
1943 n = arch_env->n_register_classes;
1944 for (i = 0; i < n; ++i) {
1946 const arch_register_class_t *cls = &arch_env->register_classes[i];
1947 for (j = 0; j < cls->n_regs; ++j) {
1948 const arch_register_t *reg = arch_register_for_index(cls, j);
1949 if (reg->type & arch_register_type_state) {
1950 ARR_APP1(arch_register_t*, stateregs, (arch_register_t *)reg);
1955 n = ARR_LEN(env->calls);
1956 n_states = ARR_LEN(stateregs);
1957 for (i = 0; i < n; ++i) {
1959 ir_node *call = env->calls[i];
1961 arity = get_irn_arity(call);
1963 /* the state reg inputs are the last n inputs of the calls */
1964 for (s = 0; s < n_states; ++s) {
1965 int inp = arity - n_states + s;
1966 const arch_register_t *reg = stateregs[s];
1967 ir_node *regnode = be_abi_reg_map_get(env->regs, reg);
1969 set_irn_n(call, inp, regnode);
1973 DEL_ARR_F(stateregs);
1977 * Create a trampoline entity for the given method.
1979 static ir_entity *create_trampoline(be_main_env_t *be, ir_entity *method)
1981 ir_type *type = get_entity_type(method);
1982 ident *old_id = get_entity_ld_ident(method);
1983 ident *id = id_mangle3("", old_id, "$stub");
1984 ir_type *parent = be->pic_trampolines_type;
1985 ir_entity *ent = new_entity(parent, old_id, type);
1986 set_entity_ld_ident(ent, id);
1987 set_entity_visibility(ent, ir_visibility_private);
1993 * Returns the trampoline entity for the given method.
1995 static ir_entity *get_trampoline(be_main_env_t *env, ir_entity *method)
1997 ir_entity *result = (ir_entity*)pmap_get(env->ent_trampoline_map, method);
1998 if (result == NULL) {
1999 result = create_trampoline(env, method);
2000 pmap_insert(env->ent_trampoline_map, method, result);
2006 static ir_entity *create_pic_symbol(be_main_env_t *be, ir_entity *entity)
2008 ident *old_id = get_entity_ld_ident(entity);
2009 ident *id = id_mangle3("", old_id, "$non_lazy_ptr");
2010 ir_type *e_type = get_entity_type(entity);
2011 ir_type *type = new_type_pointer(e_type);
2012 ir_type *parent = be->pic_symbols_type;
2013 ir_entity *ent = new_entity(parent, old_id, type);
2014 set_entity_ld_ident(ent, id);
2015 set_entity_visibility(ent, ir_visibility_private);
2020 static ir_entity *get_pic_symbol(be_main_env_t *env, ir_entity *entity)
2022 ir_entity *result = (ir_entity*)pmap_get(env->ent_pic_symbol_map, entity);
2023 if (result == NULL) {
2024 result = create_pic_symbol(env, entity);
2025 pmap_insert(env->ent_pic_symbol_map, entity, result);
2034 * Returns non-zero if a given entity can be accessed using a relative address.
2036 static int can_address_relative(ir_entity *entity)
2038 return get_entity_visibility(entity) != ir_visibility_external
2039 && !(get_entity_linkage(entity) & IR_LINKAGE_MERGE);
2042 static ir_node *get_pic_base(ir_graph *irg)
2044 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
2045 if (arch_env->impl->get_pic_base == NULL)
2047 return arch_env->impl->get_pic_base(irg);
2050 /** patches SymConsts to work in position independent code */
2051 static void fix_pic_symconsts(ir_node *node, void *data)
2053 ir_graph *irg = get_irn_irg(node);
2054 be_main_env_t *be = be_get_irg_main_env(irg);
2064 arity = get_irn_arity(node);
2065 for (i = 0; i < arity; ++i) {
2067 ir_node *pred = get_irn_n(node, i);
2069 ir_entity *pic_symbol;
2070 ir_node *pic_symconst;
2072 if (!is_SymConst(pred))
2075 entity = get_SymConst_entity(pred);
2076 block = get_nodes_block(pred);
2078 /* calls can jump to relative addresses, so we can directly jump to
2079 the (relatively) known call address or the trampoline */
2080 if (i == 1 && is_Call(node)) {
2081 ir_entity *trampoline;
2082 ir_node *trampoline_const;
2084 if (can_address_relative(entity))
2087 dbgi = get_irn_dbg_info(pred);
2088 trampoline = get_trampoline(be, entity);
2089 trampoline_const = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
2091 set_irn_n(node, i, trampoline_const);
2095 /* everything else is accessed relative to EIP */
2096 mode = get_irn_mode(pred);
2097 pic_base = get_pic_base(irg);
2099 /* all ok now for locally constructed stuff */
2100 if (can_address_relative(entity)) {
2101 ir_node *add = new_r_Add(block, pic_base, pred, mode);
2103 /* make sure the walker doesn't visit this add again */
2104 mark_irn_visited(add);
2105 set_irn_n(node, i, add);
2109 /* get entry from pic symbol segment */
2110 dbgi = get_irn_dbg_info(pred);
2111 pic_symbol = get_pic_symbol(be, entity);
2112 pic_symconst = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
2114 add = new_r_Add(block, pic_base, pic_symconst, mode);
2115 mark_irn_visited(add);
2117 /* we need an extra indirection for global data outside our current
2118 module. The loads are always safe and can therefore float
2119 and need no memory input */
2120 load = new_r_Load(block, get_irg_no_mem(irg), add, mode, cons_floats);
2121 load_res = new_r_Proj(load, mode, pn_Load_res);
2123 set_irn_n(node, i, load_res);
2127 void be_abi_introduce(ir_graph *irg)
2129 be_abi_irg_t *env = XMALLOCZ(be_abi_irg_t);
2130 ir_node *old_frame = get_irg_frame(irg);
2131 be_options_t *options = be_get_irg_options(irg);
2132 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
2133 ir_entity *entity = get_irg_entity(irg);
2134 ir_type *method_type = get_entity_type(entity);
2135 be_irg_t *birg = be_birg_from_irg(irg);
2136 struct obstack *obst = &birg->obst;
2137 ir_node *dummy = new_r_Dummy(irg,
2138 arch_env->sp->reg_class->mode);
2141 /* determine allocatable registers */
2142 assert(birg->allocatable_regs == NULL);
2143 birg->allocatable_regs = rbitset_obstack_alloc(obst, arch_env->n_registers);
2144 for (r = 0; r < arch_env->n_registers; ++r) {
2145 const arch_register_t *reg = &arch_env->registers[r];
2146 if ( !(reg->type & arch_register_type_ignore)) {
2147 rbitset_set(birg->allocatable_regs, r);
2151 /* break here if backend provides a custom API.
2152 * Note: we shouldn't have to setup any be_abi_irg_t* stuff at all,
2153 * but need more cleanup to make this work
2155 be_set_irg_abi(irg, env);
2157 be_omit_fp = options->omit_fp;
2159 env->keep_map = pmap_create();
2160 env->call = be_abi_call_new(arch_env->sp->reg_class);
2161 arch_env_get_call_abi(arch_env, method_type, env->call);
2163 env->init_sp = dummy;
2164 env->calls = NEW_ARR_F(ir_node*, 0);
2169 irg_walk_graph(irg, fix_pic_symconsts, NULL, env);
2172 /* Lower all call nodes in the IRG. */
2175 /* Process the IRG */
2178 /* fix call inputs for state registers */
2179 fix_call_state_inputs(irg);
2181 /* We don't need the keep map anymore. */
2182 pmap_destroy(env->keep_map);
2183 env->keep_map = NULL;
2185 /* calls array is not needed anymore */
2186 DEL_ARR_F(env->calls);
2189 /* reroute the stack origin of the calls to the true stack origin. */
2190 exchange(dummy, env->init_sp);
2191 exchange(old_frame, get_irg_frame(irg));
2193 pmap_destroy(env->regs);
2197 void be_abi_free(ir_graph *irg)
2199 be_abi_irg_t *env = be_get_irg_abi(irg);
2201 if (env->call != NULL)
2202 be_abi_call_free(env->call);
2203 assert(env->regs == NULL);
2206 be_set_irg_abi(irg, NULL);
2209 void be_put_allocatable_regs(const ir_graph *irg,
2210 const arch_register_class_t *cls, bitset_t *bs)
2212 be_irg_t *birg = be_birg_from_irg(irg);
2213 unsigned *allocatable_regs = birg->allocatable_regs;
2216 assert(bitset_size(bs) == cls->n_regs);
2217 bitset_clear_all(bs);
2218 for (i = 0; i < cls->n_regs; ++i) {
2219 const arch_register_t *reg = &cls->regs[i];
2220 if (rbitset_is_set(allocatable_regs, reg->global_index))
2225 unsigned be_get_n_allocatable_regs(const ir_graph *irg,
2226 const arch_register_class_t *cls)
2228 bitset_t *bs = bitset_alloca(cls->n_regs);
2229 be_put_allocatable_regs(irg, cls, bs);
2230 return bitset_popcount(bs);
2233 void be_set_allocatable_regs(const ir_graph *irg,
2234 const arch_register_class_t *cls,
2235 unsigned *raw_bitset)
2237 be_irg_t *birg = be_birg_from_irg(irg);
2238 unsigned *allocatable_regs = birg->allocatable_regs;
2241 rbitset_clear_all(raw_bitset, cls->n_regs);
2242 for (i = 0; i < cls->n_regs; ++i) {
2243 const arch_register_t *reg = &cls->regs[i];
2244 if (rbitset_is_set(allocatable_regs, reg->global_index))
2245 rbitset_set(raw_bitset, i);
2249 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_abi)
2250 void be_init_abi(void)
2252 FIRM_DBG_REGISTER(dbg, "firm.be.abi");