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 = (const be_abi_call_arg_t*)a;
131 const be_abi_call_arg_t *q = (const be_abi_call_arg_t*)b;
133 return !(p->is_res == q->is_res && p->pos == q->pos && p->callee == q->callee);
137 * Get an ABI call object argument.
139 * @param call the abi call
140 * @param is_res true for call results, false for call arguments
141 * @param pos position of the argument
142 * @param callee context type - if we are callee or caller
144 static be_abi_call_arg_t *get_call_arg(be_abi_call_t *call, int is_res, int pos, int callee)
146 be_abi_call_arg_t arg;
149 memset(&arg, 0, sizeof(arg));
154 hash = is_res * 128 + pos;
156 return (be_abi_call_arg_t*)set_find(call->params, &arg, sizeof(arg), hash);
160 * Set an ABI call object argument.
162 static void remember_call_arg(be_abi_call_arg_t *arg, be_abi_call_t *call, be_abi_context_t context)
164 unsigned hash = arg->is_res * 128 + arg->pos;
165 if (context & ABI_CONTEXT_CALLEE) {
167 set_insert(call->params, arg, sizeof(*arg), hash);
169 if (context & ABI_CONTEXT_CALLER) {
171 set_insert(call->params, arg, sizeof(*arg), hash);
175 /* Set the flags for a call. */
176 void be_abi_call_set_flags(be_abi_call_t *call, be_abi_call_flags_t flags, const be_abi_callbacks_t *cb)
182 /* Sets the number of bytes the stackframe is shrinked by the callee on return */
183 void be_abi_call_set_pop(be_abi_call_t *call, int pop)
189 /* Set register class for call address */
190 void be_abi_call_set_call_address_reg_class(be_abi_call_t *call, const arch_register_class_t *cls)
192 call->cls_addr = cls;
196 void be_abi_call_param_stack(be_abi_call_t *call, int arg_pos,
197 ir_mode *load_mode, unsigned alignment,
198 unsigned space_before, unsigned space_after,
199 be_abi_context_t context)
201 be_abi_call_arg_t arg;
202 memset(&arg, 0, sizeof(arg));
203 assert(alignment > 0 && "Alignment must be greater than 0");
205 arg.load_mode = load_mode;
206 arg.alignment = alignment;
207 arg.space_before = space_before;
208 arg.space_after = space_after;
212 remember_call_arg(&arg, call, context);
215 void be_abi_call_param_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
217 be_abi_call_arg_t arg;
218 memset(&arg, 0, sizeof(arg));
225 remember_call_arg(&arg, call, context);
228 void be_abi_call_res_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
230 be_abi_call_arg_t arg;
231 memset(&arg, 0, sizeof(arg));
238 remember_call_arg(&arg, call, context);
241 /* Get the flags of a ABI call object. */
242 be_abi_call_flags_t be_abi_call_get_flags(const be_abi_call_t *call)
248 * Constructor for a new ABI call object.
250 * @param cls_addr register class of the call address
252 * @return the new ABI call object
254 static be_abi_call_t *be_abi_call_new(const arch_register_class_t *cls_addr)
256 be_abi_call_t *call = XMALLOCZ(be_abi_call_t);
259 call->params = new_set(cmp_call_arg, 16);
261 call->cls_addr = cls_addr;
263 call->flags.bits.try_omit_fp = be_omit_fp;
269 * Destructor for an ABI call object.
271 static void be_abi_call_free(be_abi_call_t *call)
273 del_set(call->params);
278 * Initializes the frame layout from parts
280 * @param frame the stack layout that will be initialized
281 * @param args the stack argument layout type
282 * @param between the between layout type
283 * @param locals the method frame type
284 * @param stack_dir the stack direction: < 0 decreasing, > 0 increasing addresses
285 * @param param_map an array mapping method argument positions to the stack argument type
287 * @return the initialized stack layout
289 static be_stack_layout_t *stack_frame_init(be_stack_layout_t *frame, ir_type *args,
290 ir_type *between, ir_type *locals, int stack_dir,
291 ir_entity *param_map[])
293 frame->arg_type = args;
294 frame->between_type = between;
295 frame->frame_type = locals;
296 frame->initial_offset = 0;
297 frame->initial_bias = 0;
298 frame->stack_dir = stack_dir;
299 frame->order[1] = between;
300 frame->param_map = param_map;
303 frame->order[0] = args;
304 frame->order[2] = locals;
306 /* typical decreasing stack: locals have the
307 * lowest addresses, arguments the highest */
308 frame->order[0] = locals;
309 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 int 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 int stack_dir = arch_env->stack_dir;
344 const arch_register_t *sp = arch_env->sp;
345 be_abi_call_t *call = be_abi_call_new(sp->reg_class);
346 ir_mode *mach_mode = sp->reg_class->mode;
347 int no_alloc = call->flags.bits.frame_is_setup_on_call;
348 int n_res = get_method_n_ress(call_tp);
349 int do_seq = call->flags.bits.store_args_sequential && !no_alloc;
351 ir_node *res_proj = NULL;
352 int n_reg_params = 0;
353 int n_stack_params = 0;
356 const arch_register_t **states = NEW_ARR_F(const arch_register_t*, 0);
357 const arch_register_t **destroyed_regs = NEW_ARR_F(const arch_register_t*, 0);
361 int n_reg_results = 0;
362 const ir_edge_t *edge;
364 int *stack_param_idx;
365 int i, n, destroy_all_regs;
368 /* Let the isa fill out the abi description for that call node. */
369 arch_env_get_call_abi(arch_env, call_tp, call);
371 /* Insert code to put the stack arguments on the stack. */
372 assert(get_Call_n_params(irn) == n_params);
373 stack_param_idx = ALLOCAN(int, n_params);
374 for (i = 0; i < n_params; ++i) {
375 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 0);
378 int arg_size = get_type_size_bytes(get_method_param_type(call_tp, i));
380 stack_size += round_up2(arg->space_before, arg->alignment);
381 stack_size += round_up2(arg_size, arg->alignment);
382 stack_size += round_up2(arg->space_after, arg->alignment);
384 stack_param_idx[n_stack_params++] = i;
388 /* Collect all arguments which are passed in registers. */
389 reg_param_idxs = ALLOCAN(int, n_params);
390 for (i = 0; i < n_params; ++i) {
391 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 0);
392 if (arg && arg->in_reg) {
393 reg_param_idxs[n_reg_params++] = i;
398 * If the stack is decreasing and we do not want to store sequentially,
399 * or someone else allocated the call frame
400 * we allocate as much space on the stack all parameters need, by
401 * moving the stack pointer along the stack's direction.
403 * Note: we also have to do this for stack_size == 0, because we may have
404 * to adjust stack alignment for the call.
406 if (stack_dir < 0 && !do_seq && !no_alloc) {
407 curr_sp = be_new_IncSP(sp, bl, curr_sp, stack_size, 1);
410 dbgi = get_irn_dbg_info(irn);
411 /* If there are some parameters which shall be passed on the stack. */
412 if (n_stack_params > 0) {
414 ir_node **in = ALLOCAN(ir_node*, n_stack_params+1);
418 * Reverse list of stack parameters if call arguments are from left to right.
419 * We must them reverse again if they are pushed (not stored) and the stack
420 * direction is downwards.
422 if (call->flags.bits.left_to_right ^ (do_seq && stack_dir < 0)) {
423 for (i = 0; i < n_stack_params >> 1; ++i) {
424 int other = n_stack_params - i - 1;
425 int tmp = stack_param_idx[i];
426 stack_param_idx[i] = stack_param_idx[other];
427 stack_param_idx[other] = tmp;
431 curr_mem = get_Call_mem(irn);
433 in[n_in++] = curr_mem;
436 for (i = 0; i < n_stack_params; ++i) {
437 int p = stack_param_idx[i];
438 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
439 ir_node *param = get_Call_param(irn, p);
440 ir_node *addr = curr_sp;
442 ir_type *param_type = get_method_param_type(call_tp, p);
443 int param_size = get_type_size_bytes(param_type) + arg->space_after;
446 * If we wanted to build the arguments sequentially,
447 * the stack pointer for the next must be incremented,
448 * and the memory value propagated.
452 addr = curr_sp = be_new_IncSP(sp, bl, curr_sp,
453 param_size + arg->space_before, 0);
454 add_irn_dep(curr_sp, curr_mem);
456 curr_ofs += arg->space_before;
457 curr_ofs = round_up2(curr_ofs, arg->alignment);
459 /* Make the expression to compute the argument's offset. */
461 ir_mode *constmode = mach_mode;
462 if (mode_is_reference(mach_mode)) {
465 addr = new_r_Const_long(irg, constmode, curr_ofs);
466 addr = new_r_Add(bl, curr_sp, addr, mach_mode);
470 /* Insert a store for primitive arguments. */
471 if (is_atomic_type(param_type)) {
472 ir_node *mem_input = do_seq ? curr_mem : new_r_NoMem(irg);
473 ir_node *store = new_rd_Store(dbgi, bl, mem_input, addr, param, cons_none);
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,
674 arg->reg, arch_register_req_type_none);
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,
685 arch_register_req_type_none);
686 arch_set_irn_register(proj, arg->reg);
688 exchange(irn, low_call);
690 /* kill the ProjT node */
691 if (res_proj != NULL) {
695 /* Make additional projs for the caller save registers
696 and the Keep node which keeps them alive. */
701 int curr_res_proj = pn_be_Call_first_res + n_reg_results;
704 n_ins = ARR_LEN(destroyed_regs) + n_reg_results + 1;
705 in = ALLOCAN(ir_node *, n_ins);
707 /* also keep the stack pointer */
708 set_irn_link(curr_sp, (void*) sp);
711 for (i = 0; i < ARR_LEN(destroyed_regs); ++i) {
712 const arch_register_t *reg = destroyed_regs[i];
713 ir_node *proj = new_r_Proj(low_call, reg->reg_class->mode, curr_res_proj);
715 /* memorize the register in the link field. we need afterwards to set the register class of the keep correctly. */
716 be_set_constr_single_reg_out(low_call, curr_res_proj, reg,
717 arch_register_req_type_none);
718 arch_set_irn_register(proj, reg);
720 set_irn_link(proj, (void*) reg);
725 for (i = 0; i < n_reg_results; ++i) {
726 ir_node *proj = res_projs[i];
727 const arch_register_t *reg = arch_get_irn_register(proj);
728 set_irn_link(proj, (void*) reg);
733 /* create the Keep for the caller save registers */
734 keep = be_new_Keep(bl, n, in);
735 for (i = 0; i < n; ++i) {
736 const arch_register_t *reg = (const arch_register_t*)get_irn_link(in[i]);
737 be_node_set_reg_class_in(keep, i, reg->reg_class);
741 /* Clean up the stack. */
742 assert(stack_size >= call->pop);
743 stack_size -= call->pop;
745 if (stack_size > 0) {
746 ir_node *mem_proj = NULL;
748 foreach_out_edge(low_call, edge) {
749 ir_node *irn = get_edge_src_irn(edge);
750 if (is_Proj(irn) && get_Proj_proj(irn) == pn_Call_M) {
757 mem_proj = new_r_Proj(low_call, mode_M, pn_be_Call_M_regular);
758 keep_alive(mem_proj);
761 /* Clean up the stack frame or revert alignment fixes if we allocated it */
763 curr_sp = be_new_IncSP(sp, bl, curr_sp, -stack_size, 0);
766 be_abi_call_free(call);
769 DEL_ARR_F(destroyed_regs);
775 * Adjust the size of a node representing a stack alloc or free for the minimum stack alignment.
777 * @param alignment the minimum stack alignment
778 * @param size the node containing the non-aligned size
779 * @param block the block where new nodes are allocated on
780 * @param dbg debug info for new nodes
782 * @return a node representing the aligned size
784 static ir_node *adjust_alloc_size(unsigned stack_alignment, ir_node *size,
785 ir_node *block, dbg_info *dbg)
787 if (stack_alignment > 1) {
793 assert(is_po2(stack_alignment));
795 mode = get_irn_mode(size);
796 tv = new_tarval_from_long(stack_alignment-1, mode);
797 irg = get_Block_irg(block);
798 mask = new_r_Const(irg, tv);
799 size = new_rd_Add(dbg, block, size, mask, mode);
801 tv = new_tarval_from_long(-(long)stack_alignment, mode);
802 mask = new_r_Const(irg, tv);
803 size = new_rd_And(dbg, block, size, mask, mode);
809 * The alloca is transformed into a back end alloca node and connected to the stack nodes.
811 static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp)
813 ir_node *block = get_nodes_block(alloc);
814 ir_graph *irg = get_Block_irg(block);
815 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
816 ir_node *alloc_mem = NULL;
817 ir_node *alloc_res = NULL;
818 ir_type *type = get_Alloc_type(alloc);
821 const ir_edge_t *edge;
826 unsigned stack_alignment;
828 /* all non-stack Alloc nodes should already be lowered before the backend */
829 assert(get_Alloc_where(alloc) == stack_alloc);
831 foreach_out_edge(alloc, edge) {
832 ir_node *irn = get_edge_src_irn(edge);
834 assert(is_Proj(irn));
835 switch (get_Proj_proj(irn)) {
847 /* Beware: currently Alloc nodes without a result might happen,
848 only escape analysis kills them and this phase runs only for object
849 oriented source. We kill the Alloc here. */
850 if (alloc_res == NULL && alloc_mem) {
851 exchange(alloc_mem, get_Alloc_mem(alloc));
855 dbg = get_irn_dbg_info(alloc);
856 count = get_Alloc_count(alloc);
858 /* we might need to multiply the count with the element size */
859 if (type != firm_unknown_type && get_type_size_bytes(type) != 1) {
860 ir_mode *mode = get_irn_mode(count);
861 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type),
863 ir_node *cnst = new_rd_Const(dbg, irg, tv);
864 size = new_rd_Mul(dbg, block, count, cnst, mode);
869 /* The stack pointer will be modified in an unknown manner.
870 We cannot omit it. */
871 env->call->flags.bits.try_omit_fp = 0;
873 stack_alignment = 1 << arch_env->stack_alignment;
874 size = adjust_alloc_size(stack_alignment, size, block, dbg);
875 new_alloc = be_new_AddSP(arch_env->sp, block, curr_sp, size);
876 set_irn_dbg_info(new_alloc, dbg);
878 if (alloc_mem != NULL) {
882 addsp_mem = new_r_Proj(new_alloc, mode_M, pn_be_AddSP_M);
884 /* We need to sync the output mem of the AddSP with the input mem
885 edge into the alloc node. */
886 ins[0] = get_Alloc_mem(alloc);
888 sync = new_r_Sync(block, 2, ins);
890 exchange(alloc_mem, sync);
893 exchange(alloc, new_alloc);
895 /* fix projnum of alloca res */
896 set_Proj_proj(alloc_res, pn_be_AddSP_res);
898 curr_sp = new_r_Proj(new_alloc, get_irn_mode(curr_sp), pn_be_AddSP_sp);
905 * The Free is transformed into a back end free node and connected to the stack nodes.
907 static ir_node *adjust_free(be_abi_irg_t *env, ir_node *free, ir_node *curr_sp)
909 ir_node *block = get_nodes_block(free);
910 ir_graph *irg = get_irn_irg(free);
911 ir_type *type = get_Free_type(free);
912 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
913 ir_mode *sp_mode = arch_env->sp->reg_class->mode;
914 dbg_info *dbg = get_irn_dbg_info(free);
915 ir_node *subsp, *mem, *res, *size, *sync;
917 unsigned stack_alignment;
919 /* all non-stack-alloc Free nodes should already be lowered before the
921 assert(get_Free_where(free) == stack_alloc);
923 /* we might need to multiply the size with the element size */
924 if (type != firm_unknown_type && get_type_size_bytes(type) != 1) {
925 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type), mode_Iu);
926 ir_node *cnst = new_rd_Const(dbg, irg, tv);
927 ir_node *mul = new_rd_Mul(dbg, block, get_Free_size(free),
931 size = get_Free_size(free);
934 stack_alignment = 1 << arch_env->stack_alignment;
935 size = adjust_alloc_size(stack_alignment, size, block, dbg);
937 /* The stack pointer will be modified in an unknown manner.
938 We cannot omit it. */
939 env->call->flags.bits.try_omit_fp = 0;
940 subsp = be_new_SubSP(arch_env->sp, block, curr_sp, size);
941 set_irn_dbg_info(subsp, dbg);
943 mem = new_r_Proj(subsp, mode_M, pn_be_SubSP_M);
944 res = new_r_Proj(subsp, sp_mode, pn_be_SubSP_sp);
946 /* we need to sync the memory */
947 in[0] = get_Free_mem(free);
949 sync = new_r_Sync(block, 2, in);
951 /* and make the AddSP dependent on the former memory */
952 add_irn_dep(subsp, get_Free_mem(free));
955 exchange(free, sync);
962 * Check if a node is somehow data dependent on another one.
963 * both nodes must be in the same basic block.
964 * @param n1 The first node.
965 * @param n2 The second node.
966 * @return 1, if n1 is data dependent (transitively) on n2, 0 if not.
968 static int dependent_on(ir_node *n1, ir_node *n2)
970 assert(get_nodes_block(n1) == get_nodes_block(n2));
972 return heights_reachable_in_block(ir_heights, n1, n2);
975 static int cmp_call_dependency(const void *c1, const void *c2)
977 ir_node *n1 = *(ir_node **) c1;
978 ir_node *n2 = *(ir_node **) c2;
981 Classical qsort() comparison function behavior:
982 0 if both elements are equal
983 1 if second is "smaller" that first
984 -1 if first is "smaller" that second
986 if (dependent_on(n1, n2))
989 if (dependent_on(n2, n1))
992 /* The nodes have no depth order, but we need a total order because qsort()
994 return get_irn_idx(n1) - get_irn_idx(n2);
998 * Walker: links all Call/Alloc/Free nodes to the Block they are contained.
999 * Clears the irg_is_leaf flag if a Call is detected.
1001 static void link_ops_in_block_walker(ir_node *irn, void *data)
1003 be_abi_irg_t *env = (be_abi_irg_t*)data;
1004 unsigned code = get_irn_opcode(irn);
1006 if (code == iro_Call ||
1007 (code == iro_Alloc && get_Alloc_where(irn) == stack_alloc) ||
1008 (code == iro_Free && get_Free_where(irn) == stack_alloc)) {
1009 ir_node *bl = get_nodes_block(irn);
1010 void *save = get_irn_link(bl);
1012 if (code == iro_Call)
1013 env->call->flags.bits.irg_is_leaf = 0;
1015 set_irn_link(irn, save);
1016 set_irn_link(bl, irn);
1019 if (code == iro_Builtin && get_Builtin_kind(irn) == ir_bk_return_address) {
1020 ir_node *param = get_Builtin_param(irn, 0);
1021 ir_tarval *tv = get_Const_tarval(param);
1022 unsigned long value = get_tarval_long(tv);
1023 /* use ebp, so the climbframe algo works... */
1025 env->call->flags.bits.try_omit_fp = 0;
1032 * Process all Call/Alloc/Free nodes inside a basic block.
1033 * Note that the link field of the block must contain a linked list of all
1034 * Call nodes inside the Block. We first order this list according to data dependency
1035 * and that connect the calls together.
1037 static void process_ops_in_block(ir_node *bl, void *data)
1039 be_abi_irg_t *env = (be_abi_irg_t*)data;
1040 ir_node *curr_sp = env->init_sp;
1047 for (irn = (ir_node*)get_irn_link(bl); irn != NULL;
1048 irn = (ir_node*)get_irn_link(irn)) {
1052 nodes = ALLOCAN(ir_node*, n_nodes);
1053 for (irn = (ir_node*)get_irn_link(bl), n = 0; irn != NULL;
1054 irn = (ir_node*)get_irn_link(irn), ++n) {
1058 /* If there were call nodes in the block. */
1063 /* order the call nodes according to data dependency */
1064 qsort(nodes, n_nodes, sizeof(nodes[0]), cmp_call_dependency);
1066 for (i = n_nodes - 1; i >= 0; --i) {
1067 ir_node *irn = nodes[i];
1069 DBG((dbg, LEVEL_3, "\tprocessing call %+F\n", irn));
1070 switch (get_irn_opcode(irn)) {
1073 /* The stack pointer will be modified due to a call. */
1074 env->call->flags.bits.try_omit_fp = 0;
1076 curr_sp = adjust_call(env, irn, curr_sp);
1079 if (get_Alloc_where(irn) == stack_alloc)
1080 curr_sp = adjust_alloc(env, irn, curr_sp);
1083 if (get_Free_where(irn) == stack_alloc)
1084 curr_sp = adjust_free(env, irn, curr_sp);
1087 panic("invalid call");
1091 /* Keep the last stack state in the block by tying it to Keep node,
1092 * the proj from calls is already kept */
1093 if (curr_sp != env->init_sp &&
1094 !(is_Proj(curr_sp) && be_is_Call(get_Proj_pred(curr_sp)))) {
1096 keep = be_new_Keep(bl, 1, nodes);
1097 pmap_insert(env->keep_map, bl, keep);
1101 set_irn_link(bl, curr_sp);
1105 * Adjust all call nodes in the graph to the ABI conventions.
1107 static void process_calls(ir_graph *irg)
1109 be_abi_irg_t *abi = be_get_irg_abi(irg);
1111 abi->call->flags.bits.irg_is_leaf = 1;
1112 irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, abi);
1114 ir_heights = heights_new(irg);
1115 irg_block_walk_graph(irg, NULL, process_ops_in_block, abi);
1116 heights_free(ir_heights);
1120 * Computes the stack argument layout type.
1121 * Changes a possibly allocated value param type by moving
1122 * entities to the stack layout type.
1124 * @param env the ABI environment
1125 * @param call the current call ABI
1126 * @param method_type the method type
1127 * @param val_param_tp the value parameter type, will be destroyed
1128 * @param param_map an array mapping method arguments to the stack layout type
1130 * @return the stack argument layout type
1132 static ir_type *compute_arg_type(be_abi_irg_t *env, ir_graph *irg,
1133 be_abi_call_t *call,
1134 ir_type *method_type, ir_type *val_param_tp,
1135 ir_entity ***param_map)
1137 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1138 int dir = env->call->flags.bits.left_to_right ? 1 : -1;
1139 int inc = arch_env->stack_dir * dir;
1140 int n = get_method_n_params(method_type);
1141 int curr = inc > 0 ? 0 : n - 1;
1142 struct obstack *obst = be_get_be_obst(irg);
1148 ident *id = get_entity_ident(get_irg_entity(irg));
1151 *param_map = map = OALLOCN(obst, ir_entity*, n);
1152 res = new_type_struct(id_mangle_u(id, new_id_from_chars("arg_type", 8)));
1153 for (i = 0; i < n; ++i, curr += inc) {
1154 ir_type *param_type = get_method_param_type(method_type, curr);
1155 be_abi_call_arg_t *arg = get_call_arg(call, 0, curr, 1);
1158 if (arg->on_stack) {
1159 if (val_param_tp != NULL) {
1160 /* the entity was already created, create a copy in the param type */
1161 ir_entity *val_ent = get_method_value_param_ent(method_type, i);
1162 arg->stack_ent = copy_entity_own(val_ent, res);
1163 set_entity_link(val_ent, arg->stack_ent);
1164 set_entity_link(arg->stack_ent, NULL);
1166 /* create a new entity */
1167 snprintf(buf, sizeof(buf), "param_%d", i);
1168 arg->stack_ent = new_entity(res, new_id_from_str(buf), param_type);
1170 ofs += arg->space_before;
1171 ofs = round_up2(ofs, arg->alignment);
1172 set_entity_offset(arg->stack_ent, ofs);
1173 ofs += arg->space_after;
1174 ofs += get_type_size_bytes(param_type);
1175 map[i] = arg->stack_ent;
1178 set_type_size_bytes(res, ofs);
1179 set_type_state(res, layout_fixed);
1184 const arch_register_t *reg;
1188 static int cmp_regs(const void *a, const void *b)
1190 const reg_node_map_t *p = (const reg_node_map_t*)a;
1191 const reg_node_map_t *q = (const reg_node_map_t*)b;
1193 if (p->reg->reg_class == q->reg->reg_class)
1194 return p->reg->index - q->reg->index;
1196 return p->reg->reg_class - q->reg->reg_class;
1199 static void reg_map_to_arr(reg_node_map_t *res, pmap *reg_map)
1202 int n = pmap_count(reg_map);
1205 foreach_pmap(reg_map, ent) {
1206 res[i].reg = (const arch_register_t*)ent->key;
1207 res[i].irn = (ir_node*)ent->value;
1211 qsort(res, n, sizeof(res[0]), cmp_regs);
1215 * Creates a barrier.
1217 static ir_node *create_barrier(ir_node *bl, ir_node **mem, pmap *regs,
1220 int n_regs = pmap_count(regs);
1226 in = ALLOCAN(ir_node*, n_regs+1);
1227 rm = ALLOCAN(reg_node_map_t, n_regs);
1228 reg_map_to_arr(rm, regs);
1229 for (n = 0; n < n_regs; ++n) {
1237 irn = be_new_Barrier(bl, n, in);
1239 for (n = 0; n < n_regs; ++n) {
1240 ir_node *pred = rm[n].irn;
1241 const arch_register_t *reg = rm[n].reg;
1242 arch_register_req_type_t add_type = arch_register_req_type_none;
1244 const backend_info_t *info;
1246 /* stupid workaround for now... as not all nodes report register
1248 info = be_get_info(skip_Proj(pred));
1249 if (info != NULL && info->out_infos != NULL) {
1250 const arch_register_req_t *ireq = arch_get_register_req_out(pred);
1251 if (ireq->type & arch_register_req_type_ignore)
1252 add_type |= arch_register_req_type_ignore;
1253 if (ireq->type & arch_register_req_type_produces_sp)
1254 add_type |= arch_register_req_type_produces_sp;
1257 proj = new_r_Proj(irn, get_irn_mode(pred), n);
1258 be_node_set_reg_class_in(irn, n, reg->reg_class);
1260 be_set_constr_single_reg_in(irn, n, reg,
1261 arch_register_req_type_none);
1263 be_set_constr_single_reg_out(irn, n, reg, add_type);
1264 arch_set_irn_register(proj, reg);
1266 pmap_insert(regs, (void *) reg, proj);
1270 *mem = new_r_Proj(irn, mode_M, n);
1277 * Creates a be_Return for a Return node.
1279 * @param @env the abi environment
1280 * @param irn the Return node or NULL if there was none
1281 * @param bl the block where the be_Retun should be placed
1282 * @param mem the current memory
1283 * @param n_res number of return results
1285 static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl,
1286 ir_node *mem, int n_res)
1288 be_abi_call_t *call = env->call;
1289 ir_graph *irg = get_Block_irg(bl);
1290 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1292 pmap *reg_map = pmap_create();
1293 ir_node *keep = (ir_node*)pmap_get(env->keep_map, bl);
1300 const arch_register_t **regs;
1304 get the valid stack node in this block.
1305 If we had a call in that block there is a Keep constructed by process_calls()
1306 which points to the last stack modification in that block. we'll use
1307 it then. Else we use the stack from the start block and let
1308 the ssa construction fix the usage.
1310 stack = be_abi_reg_map_get(env->regs, arch_env->sp);
1312 stack = get_irn_n(keep, 0);
1314 remove_End_keepalive(get_irg_end(irg), keep);
1317 /* Insert results for Return into the register map. */
1318 for (i = 0; i < n_res; ++i) {
1319 ir_node *res = get_Return_res(irn, i);
1320 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1321 assert(arg->in_reg && "return value must be passed in register");
1322 pmap_insert(reg_map, (void *) arg->reg, res);
1325 /* Add uses of the callee save registers. */
1326 foreach_pmap(env->regs, ent) {
1327 const arch_register_t *reg = (const arch_register_t*)ent->key;
1328 if (arch_register_type_is(reg, callee_save) || arch_register_type_is(reg, ignore))
1329 pmap_insert(reg_map, ent->key, ent->value);
1332 be_abi_reg_map_set(reg_map, arch_env->sp, stack);
1334 /* Make the Epilogue node and call the arch's epilogue maker. */
1335 create_barrier(bl, &mem, reg_map, 1);
1336 call->cb->epilogue(env->cb, bl, &mem, reg_map);
1339 Maximum size of the in array for Return nodes is
1340 return args + callee save/ignore registers + memory + stack pointer
1342 in_max = pmap_count(reg_map) + n_res + 2;
1344 in = ALLOCAN(ir_node*, in_max);
1345 regs = ALLOCAN(arch_register_t const*, in_max);
1348 in[1] = be_abi_reg_map_get(reg_map, arch_env->sp);
1350 regs[1] = arch_env->sp;
1353 /* clear SP entry, since it has already been grown. */
1354 pmap_insert(reg_map, (void *) arch_env->sp, NULL);
1355 for (i = 0; i < n_res; ++i) {
1356 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1358 in[n] = be_abi_reg_map_get(reg_map, arg->reg);
1359 regs[n++] = arg->reg;
1361 /* Clear the map entry to mark the register as processed. */
1362 be_abi_reg_map_set(reg_map, arg->reg, NULL);
1365 /* grow the rest of the stuff. */
1366 foreach_pmap(reg_map, ent) {
1368 in[n] = (ir_node*)ent->value;
1369 regs[n++] = (const arch_register_t*)ent->key;
1373 /* The in array for the new back end return is now ready. */
1375 dbgi = get_irn_dbg_info(irn);
1379 /* we have to pop the shadow parameter in in case of struct returns */
1381 ret = be_new_Return(dbgi, irg, bl, n_res, pop, n, in);
1383 /* Set the register classes of the return's parameter accordingly. */
1384 for (i = 0; i < n; ++i) {
1385 if (regs[i] == NULL)
1388 be_node_set_reg_class_in(ret, i, regs[i]->reg_class);
1391 /* Free the space of the Epilog's in array and the register <-> proj map. */
1392 pmap_destroy(reg_map);
1397 typedef struct ent_pos_pair ent_pos_pair;
1398 struct ent_pos_pair {
1399 ir_entity *ent; /**< a value param entity */
1400 int pos; /**< its parameter number */
1401 ent_pos_pair *next; /**< for linking */
1404 typedef struct lower_frame_sels_env_t {
1405 ent_pos_pair *value_param_list; /**< the list of all value param entities */
1406 ir_node *frame; /**< the current frame */
1407 const arch_register_class_t *sp_class; /**< register class of the stack pointer */
1408 const arch_register_class_t *link_class; /**< register class of the link pointer */
1409 ir_type *value_tp; /**< the value type if any */
1410 ir_type *frame_tp; /**< the frame type */
1411 int static_link_pos; /**< argument number of the hidden static link */
1412 } lower_frame_sels_env_t;
1415 * Return an entity from the backend for an value param entity.
1417 * @param ent an value param type entity
1418 * @param ctx context
1420 static ir_entity *get_argument_entity(ir_entity *ent, lower_frame_sels_env_t *ctx)
1422 ir_entity *argument_ent = (ir_entity*)get_entity_link(ent);
1424 if (argument_ent == NULL) {
1425 /* we have NO argument entity yet: This is bad, as we will
1426 * need one for backing store.
1429 ir_type *frame_tp = ctx->frame_tp;
1430 unsigned offset = get_type_size_bytes(frame_tp);
1431 ir_type *tp = get_entity_type(ent);
1432 unsigned align = get_type_alignment_bytes(tp);
1434 offset += align - 1;
1435 offset &= ~(align - 1);
1437 argument_ent = copy_entity_own(ent, frame_tp);
1439 /* must be automatic to set a fixed layout */
1440 set_entity_offset(argument_ent, offset);
1441 offset += get_type_size_bytes(tp);
1443 set_type_size_bytes(frame_tp, offset);
1444 set_entity_link(ent, argument_ent);
1446 return argument_ent;
1449 * Walker: Replaces Sels of frame type and
1450 * value param type entities by FrameAddress.
1451 * Links all used entities.
1453 static void lower_frame_sels_walker(ir_node *irn, void *data)
1455 lower_frame_sels_env_t *ctx = (lower_frame_sels_env_t*)data;
1458 ir_node *ptr = get_Sel_ptr(irn);
1460 if (ptr == ctx->frame) {
1461 ir_entity *ent = get_Sel_entity(irn);
1462 ir_node *bl = get_nodes_block(irn);
1465 int is_value_param = 0;
1467 if (get_entity_owner(ent) == ctx->value_tp) {
1470 /* replace by its copy from the argument type */
1471 pos = get_struct_member_index(ctx->value_tp, ent);
1472 ent = get_argument_entity(ent, ctx);
1475 nw = be_new_FrameAddr(ctx->sp_class, bl, ctx->frame, ent);
1478 /* check, if it's a param Sel and if have not seen this entity before */
1479 if (is_value_param && get_entity_link(ent) == NULL) {
1485 ARR_APP1(ent_pos_pair, ctx->value_param_list, pair);
1487 set_entity_link(ent, ctx->value_param_list);
1494 * Check if a value parameter is transmitted as a register.
1495 * This might happen if the address of an parameter is taken which is
1496 * transmitted in registers.
1498 * Note that on some architectures this case must be handled specially
1499 * because the place of the backing store is determined by their ABI.
1501 * In the default case we move the entity to the frame type and create
1502 * a backing store into the first block.
1504 static void fix_address_of_parameter_access(be_abi_irg_t *env, ir_graph *irg,
1505 ent_pos_pair *value_param_list)
1507 be_abi_call_t *call = env->call;
1508 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1509 ent_pos_pair *entry, *new_list;
1511 int i, n = ARR_LEN(value_param_list);
1514 for (i = 0; i < n; ++i) {
1515 int pos = value_param_list[i].pos;
1516 be_abi_call_arg_t *arg = get_call_arg(call, 0, pos, 1);
1519 DBG((dbg, LEVEL_2, "\targ #%d need backing store\n", pos));
1520 value_param_list[i].next = new_list;
1521 new_list = &value_param_list[i];
1524 if (new_list != NULL) {
1525 /* ok, change the graph */
1526 ir_node *start_bl = get_irg_start_block(irg);
1527 ir_node *first_bl = get_first_block_succ(start_bl);
1528 ir_node *frame, *imem, *nmem, *store, *mem, *args;
1529 optimization_state_t state;
1532 assert(first_bl && first_bl != start_bl);
1533 /* we had already removed critical edges, so the following
1534 assertion should be always true. */
1535 assert(get_Block_n_cfgpreds(first_bl) == 1);
1537 /* now create backing stores */
1538 frame = get_irg_frame(irg);
1539 imem = get_irg_initial_mem(irg);
1541 save_optimization_state(&state);
1543 nmem = new_r_Proj(get_irg_start(irg), mode_M, pn_Start_M);
1544 restore_optimization_state(&state);
1546 /* reroute all edges to the new memory source */
1547 edges_reroute(imem, nmem, irg);
1551 args = get_irg_args(irg);
1552 for (entry = new_list; entry != NULL; entry = entry->next) {
1554 ir_type *tp = get_entity_type(entry->ent);
1555 ir_mode *mode = get_type_mode(tp);
1558 /* address for the backing store */
1559 addr = be_new_FrameAddr(arch_env->sp->reg_class, first_bl, frame, entry->ent);
1562 mem = new_r_Proj(store, mode_M, pn_Store_M);
1564 /* the backing store itself */
1565 store = new_r_Store(first_bl, mem, addr,
1566 new_r_Proj(args, mode, i), cons_none);
1568 /* the new memory Proj gets the last Proj from store */
1569 set_Proj_pred(nmem, store);
1570 set_Proj_proj(nmem, pn_Store_M);
1572 /* move all entities to the frame type */
1573 frame_tp = get_irg_frame_type(irg);
1574 offset = get_type_size_bytes(frame_tp);
1576 /* we will add new entities: set the layout to undefined */
1577 assert(get_type_state(frame_tp) == layout_fixed);
1578 set_type_state(frame_tp, layout_undefined);
1579 for (entry = new_list; entry != NULL; entry = entry->next) {
1580 ir_entity *ent = entry->ent;
1582 /* If the entity is still on the argument type, move it to the
1584 * This happens if the value_param type was build due to compound
1586 if (get_entity_owner(ent) != frame_tp) {
1587 ir_type *tp = get_entity_type(ent);
1588 unsigned align = get_type_alignment_bytes(tp);
1590 offset += align - 1;
1591 offset &= ~(align - 1);
1592 set_entity_owner(ent, frame_tp);
1593 /* must be automatic to set a fixed layout */
1594 set_entity_offset(ent, offset);
1595 offset += get_type_size_bytes(tp);
1598 set_type_size_bytes(frame_tp, offset);
1599 /* fix the layout again */
1600 set_type_state(frame_tp, layout_fixed);
1605 * The start block has no jump, instead it has an initial exec Proj.
1606 * The backend wants to handle all blocks the same way, so we replace
1607 * the out cfg edge with a real jump.
1609 static void fix_start_block(ir_graph *irg)
1611 ir_node *initial_X = get_irg_initial_exec(irg);
1612 ir_node *start_block = get_irg_start_block(irg);
1613 const ir_edge_t *edge;
1615 assert(is_Proj(initial_X));
1617 foreach_out_edge(initial_X, edge) {
1618 ir_node *block = get_edge_src_irn(edge);
1620 if (is_Anchor(block))
1622 if (block != start_block) {
1623 ir_node *jmp = new_r_Jmp(start_block);
1624 set_Block_cfgpred(block, get_edge_src_pos(edge), jmp);
1625 set_irg_initial_exec(irg, jmp);
1629 panic("Initial exec has no follow block in %+F", irg);
1633 * Update the entity of Sels to the outer value parameters.
1635 static void update_outer_frame_sels(ir_node *irn, void *env)
1637 lower_frame_sels_env_t *ctx = (lower_frame_sels_env_t*)env;
1644 ptr = get_Sel_ptr(irn);
1645 if (! is_arg_Proj(ptr))
1647 if (get_Proj_proj(ptr) != ctx->static_link_pos)
1649 ent = get_Sel_entity(irn);
1651 if (get_entity_owner(ent) == ctx->value_tp) {
1652 /* replace by its copy from the argument type */
1653 pos = get_struct_member_index(ctx->value_tp, ent);
1654 ent = get_argument_entity(ent, ctx);
1655 set_Sel_entity(irn, ent);
1657 /* check, if we have not seen this entity before */
1658 if (get_entity_link(ent) == NULL) {
1664 ARR_APP1(ent_pos_pair, ctx->value_param_list, pair);
1666 set_entity_link(ent, ctx->value_param_list);
1672 * Fix access to outer local variables.
1674 static void fix_outer_variable_access(be_abi_irg_t *env,
1675 lower_frame_sels_env_t *ctx)
1681 for (i = get_class_n_members(ctx->frame_tp) - 1; i >= 0; --i) {
1682 ir_entity *ent = get_class_member(ctx->frame_tp, i);
1684 if (! is_method_entity(ent))
1687 irg = get_entity_irg(ent);
1692 * FIXME: find the number of the static link parameter
1693 * for now we assume 0 here
1695 ctx->static_link_pos = 0;
1697 irg_walk_graph(irg, NULL, update_outer_frame_sels, ctx);
1702 * Modify the irg itself and the frame type.
1704 static void modify_irg(ir_graph *irg)
1706 be_abi_irg_t *env = be_get_irg_abi(irg);
1707 be_abi_call_t *call = env->call;
1708 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1709 const arch_register_t *sp = arch_env->sp;
1710 ir_type *method_type = get_entity_type(get_irg_entity(irg));
1711 be_irg_t *birg = be_birg_from_irg(irg);
1712 struct obstack *obst = be_get_be_obst(irg);
1713 be_stack_layout_t *stack_layout = be_get_irg_stack_layout(irg);
1716 ir_node *new_mem_proj;
1722 unsigned frame_size;
1725 const arch_register_t *fp_reg;
1726 ir_node *frame_pointer;
1730 const ir_edge_t *edge;
1731 ir_type *arg_type, *bet_type, *tp;
1732 lower_frame_sels_env_t ctx;
1733 ir_entity **param_map;
1735 DBG((dbg, LEVEL_1, "introducing abi on %+F\n", irg));
1737 /* Must fetch memory here, otherwise the start Barrier gets the wrong
1738 * memory, which leads to loops in the DAG. */
1739 old_mem = get_irg_initial_mem(irg);
1741 irp_reserve_resources(irp, IR_RESOURCE_ENTITY_LINK);
1743 /* set the links of all frame entities to NULL, we use it
1744 to detect if an entity is already linked in the value_param_list */
1745 tp = get_method_value_param_type(method_type);
1748 /* clear the links of the clone type, let the
1749 original entities point to its clones */
1750 for (i = get_struct_n_members(tp) - 1; i >= 0; --i) {
1751 ir_entity *mem = get_struct_member(tp, i);
1752 set_entity_link(mem, NULL);
1756 arg_type = compute_arg_type(env, irg, call, method_type, tp, ¶m_map);
1758 /* Convert the Sel nodes in the irg to frame addr nodes: */
1759 ctx.value_param_list = NEW_ARR_F(ent_pos_pair, 0);
1760 ctx.frame = get_irg_frame(irg);
1761 ctx.sp_class = arch_env->sp->reg_class;
1762 ctx.link_class = arch_env->link_class;
1763 ctx.frame_tp = get_irg_frame_type(irg);
1765 /* layout the stackframe now */
1766 if (get_type_state(ctx.frame_tp) == layout_undefined) {
1767 default_layout_compound_type(ctx.frame_tp);
1770 /* we will possible add new entities to the frame: set the layout to undefined */
1771 assert(get_type_state(ctx.frame_tp) == layout_fixed);
1772 set_type_state(ctx.frame_tp, layout_undefined);
1774 irg_walk_graph(irg, lower_frame_sels_walker, NULL, &ctx);
1776 /* fix the frame type layout again */
1777 set_type_state(ctx.frame_tp, layout_fixed);
1778 /* align stackframe to 4 byte */
1779 frame_size = get_type_size_bytes(ctx.frame_tp);
1780 if (frame_size % 4 != 0) {
1781 set_type_size_bytes(ctx.frame_tp, frame_size + 4 - (frame_size % 4));
1784 env->regs = pmap_create();
1786 n_params = get_method_n_params(method_type);
1787 args = OALLOCNZ(obst, ir_node*, n_params);
1790 * for inner function we must now fix access to outer frame entities.
1792 fix_outer_variable_access(env, &ctx);
1794 /* Check if a value parameter is transmitted as a register.
1795 * This might happen if the address of an parameter is taken which is
1796 * transmitted in registers.
1798 * Note that on some architectures this case must be handled specially
1799 * because the place of the backing store is determined by their ABI.
1801 * In the default case we move the entity to the frame type and create
1802 * a backing store into the first block.
1804 fix_address_of_parameter_access(env, irg, ctx.value_param_list);
1806 DEL_ARR_F(ctx.value_param_list);
1807 irp_free_resources(irp, IR_RESOURCE_ENTITY_LINK);
1809 /* Fill the argument vector */
1810 arg_tuple = get_irg_args(irg);
1811 foreach_out_edge(arg_tuple, edge) {
1812 ir_node *irn = get_edge_src_irn(edge);
1813 if (! is_Anchor(irn)) {
1814 int nr = get_Proj_proj(irn);
1816 DBG((dbg, LEVEL_2, "\treading arg: %d -> %+F\n", nr, irn));
1820 bet_type = call->cb->get_between_type(env->cb);
1821 stack_frame_init(stack_layout, arg_type, bet_type,
1822 get_irg_frame_type(irg), arch_env->stack_dir, param_map);
1823 stack_layout->sp_relative = call->flags.bits.try_omit_fp;
1825 /* Count the register params and add them to the number of Projs for the RegParams node */
1826 for (i = 0; i < n_params; ++i) {
1827 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1828 if (arg->in_reg && args[i]) {
1829 assert(arg->reg != sp && "cannot use stack pointer as parameter register");
1830 assert(i == get_Proj_proj(args[i]));
1832 /* For now, associate the register with the old Proj from Start representing that argument. */
1833 pmap_insert(env->regs, (void *) arg->reg, args[i]);
1834 DBG((dbg, LEVEL_2, "\targ #%d -> reg %s\n", i, arg->reg->name));
1838 /* Collect all callee-save registers */
1839 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
1840 const arch_register_class_t *cls = &arch_env->register_classes[i];
1841 for (j = 0; j < cls->n_regs; ++j) {
1842 const arch_register_t *reg = &cls->regs[j];
1843 if (arch_register_type_is(reg, callee_save) ||
1844 arch_register_type_is(reg, state)) {
1845 pmap_insert(env->regs, (void *) reg, NULL);
1850 /* handle start block here (place a jump in the block) */
1851 fix_start_block(irg);
1853 pmap_insert(env->regs, (void *) sp, NULL);
1854 pmap_insert(env->regs, (void *) arch_env->bp, NULL);
1855 start_bl = get_irg_start_block(irg);
1856 env->start = be_new_Start(NULL, start_bl, pmap_count(env->regs) + 1);
1859 * make proj nodes for the callee save registers.
1860 * memorize them, since Return nodes get those as inputs.
1862 * Note, that if a register corresponds to an argument, the regs map contains
1863 * the old Proj from start for that argument.
1866 rm = ALLOCAN(reg_node_map_t, pmap_count(env->regs));
1867 reg_map_to_arr(rm, env->regs);
1868 for (i = 0, n = pmap_count(env->regs); i < n; ++i) {
1869 const arch_register_t *reg = rm[i].reg;
1870 ir_mode *mode = reg->reg_class->mode;
1872 arch_register_req_type_t add_type = arch_register_req_type_none;
1876 add_type |= arch_register_req_type_produces_sp | arch_register_req_type_ignore;
1879 proj = new_r_Proj(env->start, mode, nr + 1);
1880 pmap_insert(env->regs, (void *) reg, proj);
1881 be_set_constr_single_reg_out(env->start, nr + 1, reg, add_type);
1882 arch_set_irn_register(proj, reg);
1884 DBG((dbg, LEVEL_2, "\tregister save proj #%d -> reg %s\n", nr, reg->name));
1887 /* create a new initial memory proj */
1888 assert(is_Proj(old_mem));
1889 arch_set_out_register_req(env->start, 0, arch_no_register_req);
1890 new_mem_proj = new_r_Proj(env->start, mode_M, 0);
1892 set_irg_initial_mem(irg, mem);
1894 /* Generate the Prologue */
1895 fp_reg = call->cb->prologue(env->cb, &mem, env->regs, &stack_layout->initial_bias);
1897 /* do the stack allocation BEFORE the barrier, or spill code
1898 might be added before it */
1899 env->init_sp = be_abi_reg_map_get(env->regs, sp);
1900 env->init_sp = be_new_IncSP(sp, start_bl, env->init_sp, BE_STACK_FRAME_SIZE_EXPAND, 0);
1901 be_abi_reg_map_set(env->regs, sp, env->init_sp);
1903 create_barrier(start_bl, &mem, env->regs, 0);
1905 env->init_sp = be_abi_reg_map_get(env->regs, sp);
1906 arch_set_irn_register(env->init_sp, sp);
1908 frame_pointer = be_abi_reg_map_get(env->regs, fp_reg);
1909 set_irg_frame(irg, frame_pointer);
1910 rbitset_clear(birg->allocatable_regs, fp_reg->global_index);
1912 /* rewire old mem users to new mem */
1913 exchange(old_mem, mem);
1915 /* keep the mem (for functions with an endless loop = no return) */
1918 set_irg_initial_mem(irg, mem);
1920 /* Now, introduce stack param nodes for all parameters passed on the stack */
1921 for (i = 0; i < n_params; ++i) {
1922 ir_node *arg_proj = args[i];
1923 ir_node *repl = NULL;
1925 if (arg_proj != NULL) {
1926 be_abi_call_arg_t *arg;
1927 ir_type *param_type;
1928 int nr = get_Proj_proj(arg_proj);
1931 nr = MIN(nr, n_params);
1932 arg = get_call_arg(call, 0, nr, 1);
1933 param_type = get_method_param_type(method_type, nr);
1936 repl = (ir_node*)pmap_get(env->regs, arg->reg);
1937 } else if (arg->on_stack) {
1938 ir_node *addr = be_new_FrameAddr(sp->reg_class, start_bl, frame_pointer, arg->stack_ent);
1940 /* For atomic parameters which are actually used, we create a Load node. */
1941 if (is_atomic_type(param_type) && get_irn_n_edges(args[i]) > 0) {
1942 ir_mode *mode = get_type_mode(param_type);
1943 ir_mode *load_mode = arg->load_mode;
1945 ir_node *load = new_r_Load(start_bl, new_r_NoMem(irg), addr, load_mode, cons_floats);
1946 repl = new_r_Proj(load, load_mode, pn_Load_res);
1948 if (mode != load_mode) {
1949 repl = new_r_Conv(start_bl, repl, mode);
1952 /* The stack parameter is not primitive (it is a struct or array),
1953 * we thus will create a node representing the parameter's address
1959 assert(repl != NULL);
1961 /* Beware: the mode of the register parameters is always the mode of the register class
1962 which may be wrong. Add Conv's then. */
1963 mode = get_irn_mode(args[i]);
1964 if (mode != get_irn_mode(repl)) {
1965 repl = new_r_Conv(get_nodes_block(repl), repl, mode);
1967 exchange(args[i], repl);
1971 /* the arg proj is not needed anymore now and should be only used by the anchor */
1972 assert(get_irn_n_edges(arg_tuple) == 1);
1973 kill_node(arg_tuple);
1974 set_irg_args(irg, new_r_Bad(irg));
1976 /* All Return nodes hang on the End node, so look for them there. */
1977 end = get_irg_end_block(irg);
1978 for (i = 0, n = get_Block_n_cfgpreds(end); i < n; ++i) {
1979 ir_node *irn = get_Block_cfgpred(end, i);
1981 if (is_Return(irn)) {
1982 ir_node *blk = get_nodes_block(irn);
1983 ir_node *mem = get_Return_mem(irn);
1984 ir_node *ret = create_be_return(env, irn, blk, mem, get_Return_n_ress(irn));
1989 /* if we have endless loops here, n might be <= 0. Do NOT create a be_Return then,
1990 the code is dead and will never be executed. */
1993 /** Fix the state inputs of calls that still hang on unknowns */
1994 static void fix_call_state_inputs(ir_graph *irg)
1996 be_abi_irg_t *env = be_get_irg_abi(irg);
1997 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1999 arch_register_t **stateregs = NEW_ARR_F(arch_register_t*, 0);
2001 /* Collect caller save registers */
2002 n = arch_env->n_register_classes;
2003 for (i = 0; i < n; ++i) {
2005 const arch_register_class_t *cls = &arch_env->register_classes[i];
2006 for (j = 0; j < cls->n_regs; ++j) {
2007 const arch_register_t *reg = arch_register_for_index(cls, j);
2008 if (arch_register_type_is(reg, state)) {
2009 ARR_APP1(arch_register_t*, stateregs, (arch_register_t *)reg);
2014 n = ARR_LEN(env->calls);
2015 n_states = ARR_LEN(stateregs);
2016 for (i = 0; i < n; ++i) {
2018 ir_node *call = env->calls[i];
2020 arity = get_irn_arity(call);
2022 /* the state reg inputs are the last n inputs of the calls */
2023 for (s = 0; s < n_states; ++s) {
2024 int inp = arity - n_states + s;
2025 const arch_register_t *reg = stateregs[s];
2026 ir_node *regnode = be_abi_reg_map_get(env->regs, reg);
2028 set_irn_n(call, inp, regnode);
2032 DEL_ARR_F(stateregs);
2036 * Create a trampoline entity for the given method.
2038 static ir_entity *create_trampoline(be_main_env_t *be, ir_entity *method)
2040 ir_type *type = get_entity_type(method);
2041 ident *old_id = get_entity_ld_ident(method);
2042 ident *id = id_mangle3("", old_id, "$stub");
2043 ir_type *parent = be->pic_trampolines_type;
2044 ir_entity *ent = new_entity(parent, old_id, type);
2045 set_entity_ld_ident(ent, id);
2046 set_entity_visibility(ent, ir_visibility_private);
2052 * Returns the trampoline entity for the given method.
2054 static ir_entity *get_trampoline(be_main_env_t *env, ir_entity *method)
2056 ir_entity *result = (ir_entity*)pmap_get(env->ent_trampoline_map, method);
2057 if (result == NULL) {
2058 result = create_trampoline(env, method);
2059 pmap_insert(env->ent_trampoline_map, method, result);
2065 static ir_entity *create_pic_symbol(be_main_env_t *be, ir_entity *entity)
2067 ident *old_id = get_entity_ld_ident(entity);
2068 ident *id = id_mangle3("", old_id, "$non_lazy_ptr");
2069 ir_type *e_type = get_entity_type(entity);
2070 ir_type *type = new_type_pointer(e_type);
2071 ir_type *parent = be->pic_symbols_type;
2072 ir_entity *ent = new_entity(parent, old_id, type);
2073 set_entity_ld_ident(ent, id);
2074 set_entity_visibility(ent, ir_visibility_private);
2079 static ir_entity *get_pic_symbol(be_main_env_t *env, ir_entity *entity)
2081 ir_entity *result = (ir_entity*)pmap_get(env->ent_pic_symbol_map, entity);
2082 if (result == NULL) {
2083 result = create_pic_symbol(env, entity);
2084 pmap_insert(env->ent_pic_symbol_map, entity, result);
2093 * Returns non-zero if a given entity can be accessed using a relative address.
2095 static int can_address_relative(ir_entity *entity)
2097 return get_entity_visibility(entity) != ir_visibility_external
2098 && !(get_entity_linkage(entity) & IR_LINKAGE_MERGE);
2101 static ir_node *get_pic_base(ir_graph *irg)
2103 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
2104 if (arch_env->impl->get_pic_base == NULL)
2106 return arch_env->impl->get_pic_base(irg);
2109 /** patches SymConsts to work in position independent code */
2110 static void fix_pic_symconsts(ir_node *node, void *data)
2112 ir_graph *irg = get_irn_irg(node);
2113 be_main_env_t *be = be_get_irg_main_env(irg);
2123 arity = get_irn_arity(node);
2124 for (i = 0; i < arity; ++i) {
2126 ir_node *pred = get_irn_n(node, i);
2128 ir_entity *pic_symbol;
2129 ir_node *pic_symconst;
2131 if (!is_SymConst(pred))
2134 entity = get_SymConst_entity(pred);
2135 block = get_nodes_block(pred);
2137 /* calls can jump to relative addresses, so we can directly jump to
2138 the (relatively) known call address or the trampoline */
2139 if (i == 1 && is_Call(node)) {
2140 ir_entity *trampoline;
2141 ir_node *trampoline_const;
2143 if (can_address_relative(entity))
2146 dbgi = get_irn_dbg_info(pred);
2147 trampoline = get_trampoline(be, entity);
2148 trampoline_const = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
2150 set_irn_n(node, i, trampoline_const);
2154 /* everything else is accessed relative to EIP */
2155 mode = get_irn_mode(pred);
2156 pic_base = get_pic_base(irg);
2158 /* all ok now for locally constructed stuff */
2159 if (can_address_relative(entity)) {
2160 ir_node *add = new_r_Add(block, pic_base, pred, mode);
2162 /* make sure the walker doesn't visit this add again */
2163 mark_irn_visited(add);
2164 set_irn_n(node, i, add);
2168 /* get entry from pic symbol segment */
2169 dbgi = get_irn_dbg_info(pred);
2170 pic_symbol = get_pic_symbol(be, entity);
2171 pic_symconst = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
2173 add = new_r_Add(block, pic_base, pic_symconst, mode);
2174 mark_irn_visited(add);
2176 /* we need an extra indirection for global data outside our current
2177 module. The loads are always safe and can therefore float
2178 and need no memory input */
2179 load = new_r_Load(block, new_r_NoMem(irg), add, mode, cons_floats);
2180 load_res = new_r_Proj(load, mode, pn_Load_res);
2182 set_irn_n(node, i, load_res);
2186 be_abi_irg_t *be_abi_introduce(ir_graph *irg)
2188 be_abi_irg_t *env = XMALLOCZ(be_abi_irg_t);
2189 ir_node *old_frame = get_irg_frame(irg);
2190 be_options_t *options = be_get_irg_options(irg);
2191 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
2192 ir_entity *entity = get_irg_entity(irg);
2193 ir_type *method_type = get_entity_type(entity);
2194 be_irg_t *birg = be_birg_from_irg(irg);
2195 struct obstack *obst = &birg->obst;
2201 /* determine allocatable registers */
2202 assert(birg->allocatable_regs == NULL);
2203 birg->allocatable_regs = rbitset_obstack_alloc(obst, arch_env->n_registers);
2204 for (r = 0; r < arch_env->n_registers; ++r) {
2205 const arch_register_t *reg = &arch_env->registers[r];
2206 if ( !(reg->type & arch_register_type_ignore)) {
2207 rbitset_set(birg->allocatable_regs, r);
2211 /* break here if backend provides a custom API.
2212 * Note: we shouldn't have to setup any be_abi_irg_t* stuff at all,
2213 * but need more cleanup to make this work
2215 be_set_irg_abi(irg, env);
2217 be_omit_fp = options->omit_fp;
2219 env->dce_survivor = new_survive_dce();
2220 env->keep_map = pmap_create();
2221 env->call = be_abi_call_new(arch_env->sp->reg_class);
2222 arch_env_get_call_abi(arch_env, method_type, env->call);
2224 env->init_sp = dummy = new_r_Dummy(irg, arch_env->sp->reg_class->mode);
2225 env->calls = NEW_ARR_F(ir_node*, 0);
2228 irg_walk_graph(irg, fix_pic_symconsts, NULL, env);
2231 /* Lower all call nodes in the IRG. */
2235 Beware: init backend abi call object after processing calls,
2236 otherwise some information might be not yet available.
2238 env->cb = env->call->cb->init(env->call, irg);
2240 /* Process the IRG */
2243 /* fix call inputs for state registers */
2244 fix_call_state_inputs(irg);
2246 /* We don't need the keep map anymore. */
2247 pmap_destroy(env->keep_map);
2248 env->keep_map = NULL;
2250 /* calls array is not needed anymore */
2251 DEL_ARR_F(env->calls);
2254 /* reroute the stack origin of the calls to the true stack origin. */
2255 exchange(dummy, env->init_sp);
2256 exchange(old_frame, get_irg_frame(irg));
2258 /* Make some important node pointers survive the dead node elimination. */
2259 survive_dce_register_irn(env->dce_survivor, &env->init_sp);
2260 foreach_pmap(env->regs, ent) {
2261 survive_dce_register_irn(env->dce_survivor, (ir_node **) &ent->value);
2264 env->call->cb->done(env->cb);
2269 void be_abi_free(ir_graph *irg)
2271 be_abi_irg_t *env = be_get_irg_abi(irg);
2273 if (env->call != NULL)
2274 be_abi_call_free(env->call);
2275 if (env->dce_survivor != NULL)
2276 free_survive_dce(env->dce_survivor);
2277 if (env->regs != NULL)
2278 pmap_destroy(env->regs);
2281 be_set_irg_abi(irg, NULL);
2284 void be_put_allocatable_regs(const ir_graph *irg,
2285 const arch_register_class_t *cls, bitset_t *bs)
2287 be_irg_t *birg = be_birg_from_irg(irg);
2288 unsigned *allocatable_regs = birg->allocatable_regs;
2291 assert(bitset_size(bs) == cls->n_regs);
2292 bitset_clear_all(bs);
2293 for (i = 0; i < cls->n_regs; ++i) {
2294 const arch_register_t *reg = &cls->regs[i];
2295 if (rbitset_is_set(allocatable_regs, reg->global_index))
2300 unsigned be_get_n_allocatable_regs(const ir_graph *irg,
2301 const arch_register_class_t *cls)
2303 bitset_t *bs = bitset_alloca(cls->n_regs);
2304 be_put_allocatable_regs(irg, cls, bs);
2305 return bitset_popcount(bs);
2308 void be_set_allocatable_regs(const ir_graph *irg,
2309 const arch_register_class_t *cls,
2310 unsigned *raw_bitset)
2312 be_irg_t *birg = be_birg_from_irg(irg);
2313 unsigned *allocatable_regs = birg->allocatable_regs;
2316 rbitset_clear_all(raw_bitset, cls->n_regs);
2317 for (i = 0; i < cls->n_regs; ++i) {
2318 const arch_register_t *reg = &cls->regs[i];
2319 if (rbitset_is_set(allocatable_regs, reg->global_index))
2320 rbitset_set(raw_bitset, i);
2324 ir_node *be_abi_get_callee_save_irn(be_abi_irg_t *abi, const arch_register_t *reg)
2326 assert(arch_register_type_is(reg, callee_save));
2327 assert(pmap_contains(abi->regs, (void *) reg));
2328 return (ir_node*)pmap_get(abi->regs, (void *) reg);
2331 ir_node *be_abi_get_ignore_irn(be_abi_irg_t *abi, const arch_register_t *reg)
2333 assert(arch_register_type_is(reg, ignore));
2334 assert(pmap_contains(abi->regs, (void *) reg));
2335 return (ir_node*)pmap_get(abi->regs, (void *) reg);
2338 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_abi);
2339 void be_init_abi(void)
2341 FIRM_DBG_REGISTER(dbg, "firm.be.abi");