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"
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;
369 /* Let the isa fill out the abi description for that call node. */
370 arch_env_get_call_abi(arch_env, call_tp, call);
372 /* Insert code to put the stack arguments on the stack. */
373 assert(get_Call_n_params(irn) == n_params);
374 stack_param_idx = ALLOCAN(int, n_params);
375 for (i = 0; i < n_params; ++i) {
376 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 0);
379 int arg_size = get_type_size_bytes(get_method_param_type(call_tp, i));
381 stack_size += round_up2(arg->space_before, arg->alignment);
382 stack_size += round_up2(arg_size, arg->alignment);
383 stack_size += round_up2(arg->space_after, arg->alignment);
385 stack_param_idx[n_stack_params++] = i;
389 /* Collect all arguments which are passed in registers. */
390 reg_param_idxs = ALLOCAN(int, n_params);
391 for (i = 0; i < n_params; ++i) {
392 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 0);
393 if (arg && arg->in_reg) {
394 reg_param_idxs[n_reg_params++] = i;
399 * If the stack is decreasing and we do not want to store sequentially,
400 * or someone else allocated the call frame
401 * we allocate as much space on the stack all parameters need, by
402 * moving the stack pointer along the stack's direction.
404 * Note: we also have to do this for stack_size == 0, because we may have
405 * to adjust stack alignment for the call.
407 if (stack_dir < 0 && !do_seq && !no_alloc) {
408 curr_sp = be_new_IncSP(sp, bl, curr_sp, stack_size, 1);
411 dbgi = get_irn_dbg_info(irn);
412 /* If there are some parameters which shall be passed on the stack. */
413 if (n_stack_params > 0) {
415 ir_node **in = ALLOCAN(ir_node*, n_stack_params+1);
419 * Reverse list of stack parameters if call arguments are from left to right.
420 * We must them reverse again if they are pushed (not stored) and the stack
421 * direction is downwards.
423 if (call->flags.bits.left_to_right ^ (do_seq && stack_dir < 0)) {
424 for (i = 0; i < n_stack_params >> 1; ++i) {
425 int other = n_stack_params - i - 1;
426 int tmp = stack_param_idx[i];
427 stack_param_idx[i] = stack_param_idx[other];
428 stack_param_idx[other] = tmp;
432 curr_mem = get_Call_mem(irn);
434 in[n_in++] = curr_mem;
437 for (i = 0; i < n_stack_params; ++i) {
438 int p = stack_param_idx[i];
439 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
440 ir_node *param = get_Call_param(irn, p);
441 ir_node *addr = curr_sp;
443 ir_type *param_type = get_method_param_type(call_tp, p);
444 int param_size = get_type_size_bytes(param_type) + arg->space_after;
447 * If we wanted to build the arguments sequentially,
448 * the stack pointer for the next must be incremented,
449 * and the memory value propagated.
453 addr = curr_sp = be_new_IncSP(sp, bl, curr_sp,
454 param_size + arg->space_before, 0);
455 add_irn_dep(curr_sp, curr_mem);
457 curr_ofs += arg->space_before;
458 curr_ofs = round_up2(curr_ofs, arg->alignment);
460 /* Make the expression to compute the argument's offset. */
462 ir_mode *constmode = mach_mode;
463 if (mode_is_reference(mach_mode)) {
466 addr = new_r_Const_long(irg, constmode, curr_ofs);
467 addr = new_r_Add(bl, curr_sp, addr, mach_mode);
471 /* Insert a store for primitive arguments. */
472 if (is_atomic_type(param_type)) {
473 ir_node *mem_input = do_seq ? curr_mem : new_r_NoMem(irg);
474 ir_node *store = new_rd_Store(dbgi, bl, mem_input, addr, param, cons_none);
475 mem = new_r_Proj(store, mode_M, pn_Store_M);
477 /* Make a mem copy for compound arguments. */
480 assert(mode_is_reference(get_irn_mode(param)));
481 copy = new_rd_CopyB(dbgi, bl, curr_mem, addr, param, param_type);
482 mem = new_r_Proj(copy, mode_M, pn_CopyB_M);
485 curr_ofs += param_size;
493 /* We need the sync only, if we didn't build the stores sequentially. */
495 if (n_stack_params >= 1) {
496 curr_mem = new_r_Sync(bl, n_in, in);
498 curr_mem = get_Call_mem(irn);
503 /* check for the return_twice property */
504 destroy_all_regs = 0;
505 if (is_SymConst_addr_ent(call_ptr)) {
506 ir_entity *ent = get_SymConst_entity(call_ptr);
508 if (get_entity_additional_properties(ent) & mtp_property_returns_twice)
509 destroy_all_regs = 1;
511 ir_type *call_tp = get_Call_type(irn);
513 if (get_method_additional_properties(call_tp) & mtp_property_returns_twice)
514 destroy_all_regs = 1;
517 /* Put caller save into the destroyed set and state registers in the states
519 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
521 const arch_register_class_t *cls = &arch_env->register_classes[i];
522 for (j = 0; j < cls->n_regs; ++j) {
523 const arch_register_t *reg = arch_register_for_index(cls, j);
525 /* even if destroyed all is specified, neither SP nor FP are
526 * destroyed (else bad things will happen) */
527 if (reg == arch_env->sp || reg == arch_env->bp)
530 if (reg->type & arch_register_type_state) {
531 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
532 ARR_APP1(const arch_register_t*, states, reg);
533 /* we're already in the destroyed set so no need for further
537 if (destroy_all_regs || (reg->type & arch_register_type_caller_save)) {
538 if (!(reg->type & arch_register_type_ignore)) {
539 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
545 /* search the largest result proj number */
546 res_projs = ALLOCANZ(ir_node*, n_res);
548 foreach_out_edge(irn, edge) {
549 const ir_edge_t *res_edge;
550 ir_node *irn = get_edge_src_irn(edge);
552 if (!is_Proj(irn) || get_Proj_proj(irn) != pn_Call_T_result)
555 foreach_out_edge(irn, res_edge) {
557 ir_node *res = get_edge_src_irn(res_edge);
559 assert(is_Proj(res));
561 proj = get_Proj_proj(res);
562 assert(proj < n_res);
563 assert(res_projs[proj] == NULL);
564 res_projs[proj] = res;
570 /** TODO: this is not correct for cases where return values are passed
571 * on the stack, but no known ABI does this currently...
573 n_reg_results = n_res;
576 in = ALLOCAN(ir_node*, n_reg_params + ARR_LEN(states));
578 /* make the back end call node and set its register requirements. */
579 for (i = 0; i < n_reg_params; ++i) {
580 in[n_ins++] = get_Call_param(irn, reg_param_idxs[i]);
583 /* add state registers ins */
584 for (s = 0; s < ARR_LEN(states); ++s) {
585 const arch_register_t *reg = states[s];
586 const arch_register_class_t *cls = arch_register_get_class(reg);
588 ir_node *regnode = be_abi_reg_map_get(env->regs, reg);
589 ir_fprintf(stderr, "Adding %+F\n", regnode);
591 ir_node *regnode = new_r_Unknown(irg, arch_register_class_mode(cls));
592 in[n_ins++] = regnode;
594 assert(n_ins == (int) (n_reg_params + ARR_LEN(states)));
596 /* ins collected, build the call */
597 if (env->call->flags.bits.call_has_imm && is_SymConst(call_ptr)) {
599 low_call = be_new_Call(dbgi, irg, bl, curr_mem, curr_sp, curr_sp,
600 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
601 n_ins, in, get_Call_type(irn));
602 be_Call_set_entity(low_call, get_SymConst_entity(call_ptr));
605 low_call = be_new_Call(dbgi, irg, bl, curr_mem, curr_sp, call_ptr,
606 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
607 n_ins, in, get_Call_type(irn));
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) {
623 ir_node *proj = res_projs[i];
624 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
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, be_pos_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, be_pos_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_regular);
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;
983 Classical qsort() comparison function behavior:
984 0 if both elements are equal
985 1 if second is "smaller" that first
986 -1 if first is "smaller" that second
988 if (dependent_on(n1, n2))
991 if (dependent_on(n2, n1))
994 /* The nodes have no depth order, but we need a total order because qsort()
996 return get_irn_idx(n1) - get_irn_idx(n2);
1000 * Walker: links all Call/Alloc/Free nodes to the Block they are contained.
1001 * Clears the irg_is_leaf flag if a Call is detected.
1003 static void link_ops_in_block_walker(ir_node *irn, void *data)
1005 be_abi_irg_t *env = (be_abi_irg_t*)data;
1006 unsigned code = get_irn_opcode(irn);
1008 if (code == iro_Call ||
1009 (code == iro_Alloc && get_Alloc_where(irn) == stack_alloc) ||
1010 (code == iro_Free && get_Free_where(irn) == stack_alloc)) {
1011 ir_node *bl = get_nodes_block(irn);
1012 void *save = get_irn_link(bl);
1014 if (code == iro_Call)
1015 env->call->flags.bits.irg_is_leaf = 0;
1017 set_irn_link(irn, save);
1018 set_irn_link(bl, irn);
1021 if (code == iro_Builtin && get_Builtin_kind(irn) == ir_bk_return_address) {
1022 ir_node *param = get_Builtin_param(irn, 0);
1023 ir_tarval *tv = get_Const_tarval(param);
1024 unsigned long value = get_tarval_long(tv);
1025 /* use ebp, so the climbframe algo works... */
1027 env->call->flags.bits.try_omit_fp = 0;
1034 * Process all Call/Alloc/Free nodes inside a basic block.
1035 * Note that the link field of the block must contain a linked list of all
1036 * nodes inside the Block. We first order this list according to data dependency
1037 * and that connect the nodes together.
1039 static void process_ops_in_block(ir_node *bl, void *data)
1041 be_abi_irg_t *env = (be_abi_irg_t*)data;
1042 ir_node *curr_sp = env->init_sp;
1049 for (irn = (ir_node*)get_irn_link(bl); irn != NULL;
1050 irn = (ir_node*)get_irn_link(irn)) {
1054 nodes = ALLOCAN(ir_node*, n_nodes);
1055 for (irn = (ir_node*)get_irn_link(bl), n = 0; irn != NULL;
1056 irn = (ir_node*)get_irn_link(irn), ++n) {
1060 /* If there were call nodes in the block. */
1065 /* order the call nodes according to data dependency */
1066 qsort(nodes, n_nodes, sizeof(nodes[0]), cmp_call_dependency);
1068 for (i = n_nodes - 1; i >= 0; --i) {
1069 ir_node *irn = nodes[i];
1071 DBG((dbg, LEVEL_3, "\tprocessing call %+F\n", irn));
1072 switch (get_irn_opcode(irn)) {
1075 /* The stack pointer will be modified due to a call. */
1076 env->call->flags.bits.try_omit_fp = 0;
1078 curr_sp = adjust_call(env, irn, curr_sp);
1081 if (get_Alloc_where(irn) == stack_alloc)
1082 curr_sp = adjust_alloc(env, irn, curr_sp);
1085 if (get_Free_where(irn) == stack_alloc)
1086 curr_sp = adjust_free(env, irn, curr_sp);
1089 panic("invalid call");
1093 /* Keep the last stack state in the block by tying it to Keep node,
1094 * the proj from calls is already kept */
1095 if (curr_sp != env->init_sp &&
1096 !(is_Proj(curr_sp) && be_is_Call(get_Proj_pred(curr_sp)))) {
1098 keep = be_new_Keep(bl, 1, nodes);
1099 pmap_insert(env->keep_map, bl, keep);
1103 set_irn_link(bl, curr_sp);
1107 * Adjust all call nodes in the graph to the ABI conventions.
1109 static void process_calls(ir_graph *irg)
1111 be_abi_irg_t *abi = be_get_irg_abi(irg);
1113 abi->call->flags.bits.irg_is_leaf = 1;
1114 irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, abi);
1116 ir_heights = heights_new(irg);
1117 irg_block_walk_graph(irg, NULL, process_ops_in_block, abi);
1118 heights_free(ir_heights);
1122 * Computes the stack argument layout type.
1123 * Changes a possibly allocated value param type by moving
1124 * entities to the stack layout type.
1126 * @param env the ABI environment
1127 * @param call the current call ABI
1128 * @param method_type the method type
1129 * @param val_param_tp the value parameter type, will be destroyed
1130 * @param param_map an array mapping method arguments to the stack layout type
1132 * @return the stack argument layout type
1134 static ir_type *compute_arg_type(be_abi_irg_t *env, ir_graph *irg,
1135 be_abi_call_t *call,
1136 ir_type *method_type, ir_type *val_param_tp,
1137 ir_entity ***param_map)
1139 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1140 int dir = env->call->flags.bits.left_to_right ? 1 : -1;
1141 int inc = arch_env->stack_dir * dir;
1142 int n = get_method_n_params(method_type);
1143 int curr = inc > 0 ? 0 : n - 1;
1144 struct obstack *obst = be_get_be_obst(irg);
1150 ident *id = get_entity_ident(get_irg_entity(irg));
1153 *param_map = map = OALLOCN(obst, ir_entity*, n);
1154 res = new_type_struct(id_mangle_u(id, new_id_from_chars("arg_type", 8)));
1155 for (i = 0; i < n; ++i, curr += inc) {
1156 ir_type *param_type = get_method_param_type(method_type, curr);
1157 be_abi_call_arg_t *arg = get_call_arg(call, 0, curr, 1);
1160 if (arg->on_stack) {
1161 if (val_param_tp != NULL) {
1162 /* the entity was already created, create a copy in the param type */
1163 ir_entity *val_ent = get_method_value_param_ent(method_type, i);
1164 arg->stack_ent = copy_entity_own(val_ent, res);
1165 set_entity_link(val_ent, arg->stack_ent);
1166 set_entity_link(arg->stack_ent, NULL);
1168 /* create a new entity */
1169 snprintf(buf, sizeof(buf), "param_%d", i);
1170 arg->stack_ent = new_entity(res, new_id_from_str(buf), param_type);
1172 ofs += arg->space_before;
1173 ofs = round_up2(ofs, arg->alignment);
1174 set_entity_offset(arg->stack_ent, ofs);
1175 ofs += arg->space_after;
1176 ofs += get_type_size_bytes(param_type);
1177 map[i] = arg->stack_ent;
1180 set_type_size_bytes(res, ofs);
1181 set_type_state(res, layout_fixed);
1186 const arch_register_t *reg;
1190 static int cmp_regs(const void *a, const void *b)
1192 const reg_node_map_t *p = (const reg_node_map_t*)a;
1193 const reg_node_map_t *q = (const reg_node_map_t*)b;
1195 if (p->reg->reg_class == q->reg->reg_class)
1196 return p->reg->index - q->reg->index;
1198 return p->reg->reg_class < q->reg->reg_class ? -1 : +1;
1201 static void reg_map_to_arr(reg_node_map_t *res, pmap *reg_map)
1204 size_t n = pmap_count(reg_map);
1207 foreach_pmap(reg_map, ent) {
1208 res[i].reg = (const arch_register_t*)ent->key;
1209 res[i].irn = (ir_node*)ent->value;
1213 qsort(res, n, sizeof(res[0]), cmp_regs);
1217 * Creates a barrier.
1219 static ir_node *create_barrier(ir_node *bl, ir_node **mem, pmap *regs,
1222 int n_regs = pmap_count(regs);
1228 in = ALLOCAN(ir_node*, n_regs+1);
1229 rm = ALLOCAN(reg_node_map_t, n_regs);
1230 reg_map_to_arr(rm, regs);
1231 for (n = 0; n < n_regs; ++n) {
1239 irn = be_new_Barrier(bl, n, in);
1241 for (n = 0; n < n_regs; ++n) {
1242 ir_node *pred = rm[n].irn;
1243 const arch_register_t *reg = rm[n].reg;
1244 arch_register_req_type_t add_type = arch_register_req_type_none;
1246 const backend_info_t *info;
1248 /* stupid workaround for now... as not all nodes report register
1250 info = be_get_info(skip_Proj(pred));
1251 if (info != NULL && info->out_infos != NULL) {
1252 const arch_register_req_t *ireq = arch_get_register_req_out(pred);
1253 if (ireq->type & arch_register_req_type_ignore)
1254 add_type |= arch_register_req_type_ignore;
1255 if (ireq->type & arch_register_req_type_produces_sp)
1256 add_type |= arch_register_req_type_produces_sp;
1259 proj = new_r_Proj(irn, get_irn_mode(pred), n);
1260 be_node_set_reg_class_in(irn, n, reg->reg_class);
1262 be_set_constr_single_reg_in(irn, n, reg,
1263 arch_register_req_type_none);
1265 be_set_constr_single_reg_out(irn, n, reg, add_type);
1266 arch_set_irn_register(proj, reg);
1268 pmap_insert(regs, (void *) reg, proj);
1272 *mem = new_r_Proj(irn, mode_M, n);
1279 * Creates a be_Return for a Return node.
1281 * @param @env the abi environment
1282 * @param irn the Return node or NULL if there was none
1283 * @param bl the block where the be_Retun should be placed
1284 * @param mem the current memory
1285 * @param n_res number of return results
1287 static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl,
1288 ir_node *mem, int n_res)
1290 be_abi_call_t *call = env->call;
1291 ir_graph *irg = get_Block_irg(bl);
1292 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1294 pmap *reg_map = pmap_create();
1295 ir_node *keep = (ir_node*)pmap_get(env->keep_map, bl);
1302 const arch_register_t **regs;
1306 get the valid stack node in this block.
1307 If we had a call in that block there is a Keep constructed by process_calls()
1308 which points to the last stack modification in that block. we'll use
1309 it then. Else we use the stack from the start block and let
1310 the ssa construction fix the usage.
1312 stack = be_abi_reg_map_get(env->regs, arch_env->sp);
1314 stack = get_irn_n(keep, 0);
1316 remove_End_keepalive(get_irg_end(irg), keep);
1319 /* Insert results for Return into the register map. */
1320 for (i = 0; i < n_res; ++i) {
1321 ir_node *res = get_Return_res(irn, i);
1322 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1323 assert(arg->in_reg && "return value must be passed in register");
1324 pmap_insert(reg_map, (void *) arg->reg, res);
1327 /* Add uses of the callee save registers. */
1328 foreach_pmap(env->regs, ent) {
1329 const arch_register_t *reg = (const arch_register_t*)ent->key;
1330 if (reg->type & (arch_register_type_callee_save | arch_register_type_ignore))
1331 pmap_insert(reg_map, ent->key, ent->value);
1334 be_abi_reg_map_set(reg_map, arch_env->sp, stack);
1336 /* Make the Epilogue node and call the arch's epilogue maker. */
1337 create_barrier(bl, &mem, reg_map, 1);
1338 call->cb->epilogue(env->cb, bl, &mem, reg_map);
1341 Maximum size of the in array for Return nodes is
1342 return args + callee save/ignore registers + memory + stack pointer
1344 in_max = pmap_count(reg_map) + n_res + 2;
1346 in = ALLOCAN(ir_node*, in_max);
1347 regs = ALLOCAN(arch_register_t const*, in_max);
1350 in[1] = be_abi_reg_map_get(reg_map, arch_env->sp);
1352 regs[1] = arch_env->sp;
1355 /* clear SP entry, since it has already been grown. */
1356 pmap_insert(reg_map, (void *) arch_env->sp, NULL);
1357 for (i = 0; i < n_res; ++i) {
1358 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1360 in[n] = be_abi_reg_map_get(reg_map, arg->reg);
1361 regs[n++] = arg->reg;
1363 /* Clear the map entry to mark the register as processed. */
1364 be_abi_reg_map_set(reg_map, arg->reg, NULL);
1367 /* grow the rest of the stuff. */
1368 foreach_pmap(reg_map, ent) {
1370 in[n] = (ir_node*)ent->value;
1371 regs[n++] = (const arch_register_t*)ent->key;
1375 /* The in array for the new back end return is now ready. */
1377 dbgi = get_irn_dbg_info(irn);
1381 /* we have to pop the shadow parameter in in case of struct returns */
1383 ret = be_new_Return(dbgi, irg, bl, n_res, pop, n, in);
1385 /* Set the register classes of the return's parameter accordingly. */
1386 for (i = 0; i < n; ++i) {
1387 if (regs[i] == NULL)
1390 be_node_set_reg_class_in(ret, i, regs[i]->reg_class);
1393 /* Free the space of the Epilog's in array and the register <-> proj map. */
1394 pmap_destroy(reg_map);
1399 typedef struct ent_pos_pair ent_pos_pair;
1400 struct ent_pos_pair {
1401 ir_entity *ent; /**< a value param entity */
1402 int pos; /**< its parameter number */
1403 ent_pos_pair *next; /**< for linking */
1406 typedef struct lower_frame_sels_env_t {
1407 ent_pos_pair *value_param_list; /**< the list of all value param entities */
1408 ir_node *frame; /**< the current frame */
1409 const arch_register_class_t *sp_class; /**< register class of the stack pointer */
1410 const arch_register_class_t *link_class; /**< register class of the link pointer */
1411 ir_type *value_tp; /**< the value type if any */
1412 ir_type *frame_tp; /**< the frame type */
1413 int static_link_pos; /**< argument number of the hidden static link */
1414 } lower_frame_sels_env_t;
1417 * Return an entity from the backend for an value param entity.
1419 * @param ent an value param type entity
1420 * @param ctx context
1422 static ir_entity *get_argument_entity(ir_entity *ent, lower_frame_sels_env_t *ctx)
1424 ir_entity *argument_ent = (ir_entity*)get_entity_link(ent);
1426 if (argument_ent == NULL) {
1427 /* we have NO argument entity yet: This is bad, as we will
1428 * need one for backing store.
1431 ir_type *frame_tp = ctx->frame_tp;
1432 unsigned offset = get_type_size_bytes(frame_tp);
1433 ir_type *tp = get_entity_type(ent);
1434 unsigned align = get_type_alignment_bytes(tp);
1436 offset += align - 1;
1437 offset &= ~(align - 1);
1439 argument_ent = copy_entity_own(ent, frame_tp);
1441 /* must be automatic to set a fixed layout */
1442 set_entity_offset(argument_ent, offset);
1443 offset += get_type_size_bytes(tp);
1445 set_type_size_bytes(frame_tp, offset);
1446 set_entity_link(ent, argument_ent);
1448 return argument_ent;
1451 * Walker: Replaces Sels of frame type and
1452 * value param type entities by FrameAddress.
1453 * Links all used entities.
1455 static void lower_frame_sels_walker(ir_node *irn, void *data)
1457 lower_frame_sels_env_t *ctx = (lower_frame_sels_env_t*)data;
1460 ir_node *ptr = get_Sel_ptr(irn);
1462 if (ptr == ctx->frame) {
1463 ir_entity *ent = get_Sel_entity(irn);
1464 ir_node *bl = get_nodes_block(irn);
1467 int is_value_param = 0;
1469 if (get_entity_owner(ent) == ctx->value_tp) {
1472 /* replace by its copy from the argument type */
1473 pos = get_struct_member_index(ctx->value_tp, ent);
1474 ent = get_argument_entity(ent, ctx);
1477 nw = be_new_FrameAddr(ctx->sp_class, bl, ctx->frame, ent);
1480 /* check, if it's a param Sel and if have not seen this entity before */
1481 if (is_value_param && get_entity_link(ent) == NULL) {
1487 ARR_APP1(ent_pos_pair, ctx->value_param_list, pair);
1489 set_entity_link(ent, ctx->value_param_list);
1496 * Check if a value parameter is transmitted as a register.
1497 * This might happen if the address of an parameter is taken which is
1498 * transmitted in registers.
1500 * Note that on some architectures this case must be handled specially
1501 * because the place of the backing store is determined by their ABI.
1503 * In the default case we move the entity to the frame type and create
1504 * a backing store into the first block.
1506 static void fix_address_of_parameter_access(be_abi_irg_t *env, ir_graph *irg,
1507 ent_pos_pair *value_param_list)
1509 be_abi_call_t *call = env->call;
1510 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1511 ent_pos_pair *entry, *new_list;
1513 int i, n = ARR_LEN(value_param_list);
1516 for (i = 0; i < n; ++i) {
1517 int pos = value_param_list[i].pos;
1518 be_abi_call_arg_t *arg = get_call_arg(call, 0, pos, 1);
1521 DBG((dbg, LEVEL_2, "\targ #%d need backing store\n", pos));
1522 value_param_list[i].next = new_list;
1523 new_list = &value_param_list[i];
1526 if (new_list != NULL) {
1527 /* ok, change the graph */
1528 ir_node *start_bl = get_irg_start_block(irg);
1529 ir_node *first_bl = get_first_block_succ(start_bl);
1530 ir_node *frame, *imem, *nmem, *store, *mem, *args;
1531 optimization_state_t state;
1534 assert(first_bl && first_bl != start_bl);
1535 /* we had already removed critical edges, so the following
1536 assertion should be always true. */
1537 assert(get_Block_n_cfgpreds(first_bl) == 1);
1539 /* now create backing stores */
1540 frame = get_irg_frame(irg);
1541 imem = get_irg_initial_mem(irg);
1543 save_optimization_state(&state);
1545 nmem = new_r_Proj(get_irg_start(irg), mode_M, pn_Start_M);
1546 restore_optimization_state(&state);
1548 /* reroute all edges to the new memory source */
1549 edges_reroute(imem, nmem, irg);
1553 args = get_irg_args(irg);
1554 for (entry = new_list; entry != NULL; entry = entry->next) {
1556 ir_type *tp = get_entity_type(entry->ent);
1557 ir_mode *mode = get_type_mode(tp);
1560 /* address for the backing store */
1561 addr = be_new_FrameAddr(arch_env->sp->reg_class, first_bl, frame, entry->ent);
1564 mem = new_r_Proj(store, mode_M, pn_Store_M);
1566 /* the backing store itself */
1567 store = new_r_Store(first_bl, mem, addr,
1568 new_r_Proj(args, mode, i), cons_none);
1570 /* the new memory Proj gets the last Proj from store */
1571 set_Proj_pred(nmem, store);
1572 set_Proj_proj(nmem, pn_Store_M);
1574 /* move all entities to the frame type */
1575 frame_tp = get_irg_frame_type(irg);
1576 offset = get_type_size_bytes(frame_tp);
1578 /* we will add new entities: set the layout to undefined */
1579 assert(get_type_state(frame_tp) == layout_fixed);
1580 set_type_state(frame_tp, layout_undefined);
1581 for (entry = new_list; entry != NULL; entry = entry->next) {
1582 ir_entity *ent = entry->ent;
1584 /* If the entity is still on the argument type, move it to the
1586 * This happens if the value_param type was build due to compound
1588 if (get_entity_owner(ent) != frame_tp) {
1589 ir_type *tp = get_entity_type(ent);
1590 unsigned align = get_type_alignment_bytes(tp);
1592 offset += align - 1;
1593 offset &= ~(align - 1);
1594 set_entity_owner(ent, frame_tp);
1595 /* must be automatic to set a fixed layout */
1596 set_entity_offset(ent, offset);
1597 offset += get_type_size_bytes(tp);
1600 set_type_size_bytes(frame_tp, offset);
1601 /* fix the layout again */
1602 set_type_state(frame_tp, layout_fixed);
1607 * The start block has no jump, instead it has an initial exec Proj.
1608 * The backend wants to handle all blocks the same way, so we replace
1609 * the out cfg edge with a real jump.
1611 static void fix_start_block(ir_graph *irg)
1613 ir_node *initial_X = get_irg_initial_exec(irg);
1614 ir_node *start_block = get_irg_start_block(irg);
1615 const ir_edge_t *edge;
1617 assert(is_Proj(initial_X));
1619 foreach_out_edge(initial_X, edge) {
1620 ir_node *block = get_edge_src_irn(edge);
1622 if (is_Anchor(block))
1624 if (block != start_block) {
1625 ir_node *jmp = new_r_Jmp(start_block);
1626 set_Block_cfgpred(block, get_edge_src_pos(edge), jmp);
1627 set_irg_initial_exec(irg, jmp);
1631 panic("Initial exec has no follow block in %+F", irg);
1635 * Update the entity of Sels to the outer value parameters.
1637 static void update_outer_frame_sels(ir_node *irn, void *env)
1639 lower_frame_sels_env_t *ctx = (lower_frame_sels_env_t*)env;
1646 ptr = get_Sel_ptr(irn);
1647 if (! is_arg_Proj(ptr))
1649 if (get_Proj_proj(ptr) != ctx->static_link_pos)
1651 ent = get_Sel_entity(irn);
1653 if (get_entity_owner(ent) == ctx->value_tp) {
1654 /* replace by its copy from the argument type */
1655 pos = get_struct_member_index(ctx->value_tp, ent);
1656 ent = get_argument_entity(ent, ctx);
1657 set_Sel_entity(irn, ent);
1659 /* check, if we have not seen this entity before */
1660 if (get_entity_link(ent) == NULL) {
1666 ARR_APP1(ent_pos_pair, ctx->value_param_list, pair);
1668 set_entity_link(ent, ctx->value_param_list);
1674 * Fix access to outer local variables.
1676 static void fix_outer_variable_access(be_abi_irg_t *env,
1677 lower_frame_sels_env_t *ctx)
1683 for (i = get_class_n_members(ctx->frame_tp) - 1; i >= 0; --i) {
1684 ir_entity *ent = get_class_member(ctx->frame_tp, i);
1686 if (! is_method_entity(ent))
1689 irg = get_entity_irg(ent);
1694 * FIXME: find the number of the static link parameter
1695 * for now we assume 0 here
1697 ctx->static_link_pos = 0;
1699 irg_walk_graph(irg, NULL, update_outer_frame_sels, ctx);
1704 * Modify the irg itself and the frame type.
1706 static void modify_irg(ir_graph *irg)
1708 be_abi_irg_t *env = be_get_irg_abi(irg);
1709 be_abi_call_t *call = env->call;
1710 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1711 const arch_register_t *sp = arch_env->sp;
1712 ir_type *method_type = get_entity_type(get_irg_entity(irg));
1713 be_irg_t *birg = be_birg_from_irg(irg);
1714 struct obstack *obst = be_get_be_obst(irg);
1715 be_stack_layout_t *stack_layout = be_get_irg_stack_layout(irg);
1718 ir_node *new_mem_proj;
1724 unsigned frame_size;
1727 const arch_register_t *fp_reg;
1728 ir_node *frame_pointer;
1732 const ir_edge_t *edge;
1733 ir_type *arg_type, *bet_type, *tp;
1734 lower_frame_sels_env_t ctx;
1735 ir_entity **param_map;
1737 DBG((dbg, LEVEL_1, "introducing abi on %+F\n", irg));
1739 /* Must fetch memory here, otherwise the start Barrier gets the wrong
1740 * memory, which leads to loops in the DAG. */
1741 old_mem = get_irg_initial_mem(irg);
1743 irp_reserve_resources(irp, IR_RESOURCE_ENTITY_LINK);
1745 /* set the links of all frame entities to NULL, we use it
1746 to detect if an entity is already linked in the value_param_list */
1747 tp = get_method_value_param_type(method_type);
1750 /* clear the links of the clone type, let the
1751 original entities point to its clones */
1752 for (i = get_struct_n_members(tp) - 1; i >= 0; --i) {
1753 ir_entity *mem = get_struct_member(tp, i);
1754 set_entity_link(mem, NULL);
1758 arg_type = compute_arg_type(env, irg, call, method_type, tp, ¶m_map);
1760 /* Convert the Sel nodes in the irg to frame addr nodes: */
1761 ctx.value_param_list = NEW_ARR_F(ent_pos_pair, 0);
1762 ctx.frame = get_irg_frame(irg);
1763 ctx.sp_class = arch_env->sp->reg_class;
1764 ctx.link_class = arch_env->link_class;
1765 ctx.frame_tp = get_irg_frame_type(irg);
1767 /* layout the stackframe now */
1768 if (get_type_state(ctx.frame_tp) == layout_undefined) {
1769 default_layout_compound_type(ctx.frame_tp);
1772 /* we will possible add new entities to the frame: set the layout to undefined */
1773 assert(get_type_state(ctx.frame_tp) == layout_fixed);
1774 set_type_state(ctx.frame_tp, layout_undefined);
1776 irg_walk_graph(irg, lower_frame_sels_walker, NULL, &ctx);
1778 /* fix the frame type layout again */
1779 set_type_state(ctx.frame_tp, layout_fixed);
1780 /* align stackframe to 4 byte */
1781 frame_size = get_type_size_bytes(ctx.frame_tp);
1782 if (frame_size % 4 != 0) {
1783 set_type_size_bytes(ctx.frame_tp, frame_size + 4 - (frame_size % 4));
1786 env->regs = pmap_create();
1788 n_params = get_method_n_params(method_type);
1789 args = OALLOCNZ(obst, ir_node*, n_params);
1792 * for inner function we must now fix access to outer frame entities.
1794 fix_outer_variable_access(env, &ctx);
1796 /* Check if a value parameter is transmitted as a register.
1797 * This might happen if the address of an parameter is taken which is
1798 * transmitted in registers.
1800 * Note that on some architectures this case must be handled specially
1801 * because the place of the backing store is determined by their ABI.
1803 * In the default case we move the entity to the frame type and create
1804 * a backing store into the first block.
1806 fix_address_of_parameter_access(env, irg, ctx.value_param_list);
1808 DEL_ARR_F(ctx.value_param_list);
1809 irp_free_resources(irp, IR_RESOURCE_ENTITY_LINK);
1811 /* Fill the argument vector */
1812 arg_tuple = get_irg_args(irg);
1813 foreach_out_edge(arg_tuple, edge) {
1814 ir_node *irn = get_edge_src_irn(edge);
1815 if (! is_Anchor(irn)) {
1816 int nr = get_Proj_proj(irn);
1818 DBG((dbg, LEVEL_2, "\treading arg: %d -> %+F\n", nr, irn));
1822 bet_type = call->cb->get_between_type(env->cb);
1823 stack_frame_init(stack_layout, arg_type, bet_type,
1824 get_irg_frame_type(irg), arch_env->stack_dir, param_map);
1825 stack_layout->sp_relative = call->flags.bits.try_omit_fp;
1827 /* Count the register params and add them to the number of Projs for the RegParams node */
1828 for (i = 0; i < n_params; ++i) {
1829 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1830 if (arg->in_reg && args[i]) {
1831 assert(arg->reg != sp && "cannot use stack pointer as parameter register");
1832 assert(i == get_Proj_proj(args[i]));
1834 /* For now, associate the register with the old Proj from Start representing that argument. */
1835 pmap_insert(env->regs, (void *) arg->reg, args[i]);
1836 DBG((dbg, LEVEL_2, "\targ #%d -> reg %s\n", i, arg->reg->name));
1840 /* Collect all callee-save registers */
1841 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
1842 const arch_register_class_t *cls = &arch_env->register_classes[i];
1843 for (j = 0; j < cls->n_regs; ++j) {
1844 const arch_register_t *reg = &cls->regs[j];
1845 if (reg->type & (arch_register_type_callee_save | arch_register_type_state)) {
1846 pmap_insert(env->regs, (void *) reg, NULL);
1851 /* handle start block here (place a jump in the block) */
1852 fix_start_block(irg);
1854 pmap_insert(env->regs, (void *) sp, NULL);
1855 pmap_insert(env->regs, (void *) arch_env->bp, NULL);
1856 start_bl = get_irg_start_block(irg);
1857 env->start = be_new_Start(NULL, start_bl, pmap_count(env->regs) + 1);
1860 * make proj nodes for the callee save registers.
1861 * memorize them, since Return nodes get those as inputs.
1863 * Note, that if a register corresponds to an argument, the regs map contains
1864 * the old Proj from start for that argument.
1867 rm = ALLOCAN(reg_node_map_t, pmap_count(env->regs));
1868 reg_map_to_arr(rm, env->regs);
1869 for (i = 0, n = pmap_count(env->regs); i < n; ++i) {
1870 const arch_register_t *reg = rm[i].reg;
1871 ir_mode *mode = reg->reg_class->mode;
1873 arch_register_req_type_t add_type = arch_register_req_type_none;
1877 add_type |= arch_register_req_type_produces_sp | arch_register_req_type_ignore;
1880 proj = new_r_Proj(env->start, mode, nr + 1);
1881 pmap_insert(env->regs, (void *) reg, proj);
1882 be_set_constr_single_reg_out(env->start, nr + 1, reg, add_type);
1883 arch_set_irn_register(proj, reg);
1885 DBG((dbg, LEVEL_2, "\tregister save proj #%d -> reg %s\n", nr, reg->name));
1888 /* create a new initial memory proj */
1889 assert(is_Proj(old_mem));
1890 arch_set_out_register_req(env->start, 0, arch_no_register_req);
1891 new_mem_proj = new_r_Proj(env->start, mode_M, 0);
1893 set_irg_initial_mem(irg, mem);
1895 /* Generate the Prologue */
1896 fp_reg = call->cb->prologue(env->cb, &mem, env->regs, &stack_layout->initial_bias);
1898 /* do the stack allocation BEFORE the barrier, or spill code
1899 might be added before it */
1900 env->init_sp = be_abi_reg_map_get(env->regs, sp);
1901 env->init_sp = be_new_IncSP(sp, start_bl, env->init_sp, BE_STACK_FRAME_SIZE_EXPAND, 0);
1902 be_abi_reg_map_set(env->regs, sp, env->init_sp);
1904 create_barrier(start_bl, &mem, env->regs, 0);
1906 env->init_sp = be_abi_reg_map_get(env->regs, sp);
1907 arch_set_irn_register(env->init_sp, sp);
1909 frame_pointer = be_abi_reg_map_get(env->regs, fp_reg);
1910 set_irg_frame(irg, frame_pointer);
1911 rbitset_clear(birg->allocatable_regs, fp_reg->global_index);
1913 /* rewire old mem users to new mem */
1914 exchange(old_mem, mem);
1916 /* keep the mem (for functions with an endless loop = no return) */
1919 set_irg_initial_mem(irg, mem);
1921 /* Now, introduce stack param nodes for all parameters passed on the stack */
1922 for (i = 0; i < n_params; ++i) {
1923 ir_node *arg_proj = args[i];
1924 ir_node *repl = NULL;
1926 if (arg_proj != NULL) {
1927 be_abi_call_arg_t *arg;
1928 ir_type *param_type;
1929 int nr = get_Proj_proj(arg_proj);
1932 nr = MIN(nr, n_params);
1933 arg = get_call_arg(call, 0, nr, 1);
1934 param_type = get_method_param_type(method_type, nr);
1937 repl = (ir_node*)pmap_get(env->regs, arg->reg);
1938 } else if (arg->on_stack) {
1939 ir_node *addr = be_new_FrameAddr(sp->reg_class, start_bl, frame_pointer, arg->stack_ent);
1941 /* For atomic parameters which are actually used, we create a Load node. */
1942 if (is_atomic_type(param_type) && get_irn_n_edges(args[i]) > 0) {
1943 ir_mode *mode = get_type_mode(param_type);
1944 ir_mode *load_mode = arg->load_mode;
1946 ir_node *load = new_r_Load(start_bl, new_r_NoMem(irg), addr, load_mode, cons_floats);
1947 repl = new_r_Proj(load, load_mode, pn_Load_res);
1949 if (mode != load_mode) {
1950 repl = new_r_Conv(start_bl, repl, mode);
1953 /* The stack parameter is not primitive (it is a struct or array),
1954 * we thus will create a node representing the parameter's address
1960 assert(repl != NULL);
1962 /* Beware: the mode of the register parameters is always the mode of the register class
1963 which may be wrong. Add Conv's then. */
1964 mode = get_irn_mode(args[i]);
1965 if (mode != get_irn_mode(repl)) {
1966 repl = new_r_Conv(get_nodes_block(repl), repl, mode);
1968 exchange(args[i], repl);
1972 /* the arg proj is not needed anymore now and should be only used by the anchor */
1973 assert(get_irn_n_edges(arg_tuple) == 1);
1974 kill_node(arg_tuple);
1975 set_irg_args(irg, new_r_Bad(irg));
1977 /* All Return nodes hang on the End node, so look for them there. */
1978 end = get_irg_end_block(irg);
1979 for (i = 0, n = get_Block_n_cfgpreds(end); i < n; ++i) {
1980 ir_node *irn = get_Block_cfgpred(end, i);
1982 if (is_Return(irn)) {
1983 ir_node *blk = get_nodes_block(irn);
1984 ir_node *mem = get_Return_mem(irn);
1985 ir_node *ret = create_be_return(env, irn, blk, mem, get_Return_n_ress(irn));
1990 /* if we have endless loops here, n might be <= 0. Do NOT create a be_Return then,
1991 the code is dead and will never be executed. */
1994 /** Fix the state inputs of calls that still hang on unknowns */
1995 static void fix_call_state_inputs(ir_graph *irg)
1997 be_abi_irg_t *env = be_get_irg_abi(irg);
1998 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
2000 arch_register_t **stateregs = NEW_ARR_F(arch_register_t*, 0);
2002 /* Collect caller save registers */
2003 n = arch_env->n_register_classes;
2004 for (i = 0; i < n; ++i) {
2006 const arch_register_class_t *cls = &arch_env->register_classes[i];
2007 for (j = 0; j < cls->n_regs; ++j) {
2008 const arch_register_t *reg = arch_register_for_index(cls, j);
2009 if (reg->type & arch_register_type_state) {
2010 ARR_APP1(arch_register_t*, stateregs, (arch_register_t *)reg);
2015 n = ARR_LEN(env->calls);
2016 n_states = ARR_LEN(stateregs);
2017 for (i = 0; i < n; ++i) {
2019 ir_node *call = env->calls[i];
2021 arity = get_irn_arity(call);
2023 /* the state reg inputs are the last n inputs of the calls */
2024 for (s = 0; s < n_states; ++s) {
2025 int inp = arity - n_states + s;
2026 const arch_register_t *reg = stateregs[s];
2027 ir_node *regnode = be_abi_reg_map_get(env->regs, reg);
2029 set_irn_n(call, inp, regnode);
2033 DEL_ARR_F(stateregs);
2037 * Create a trampoline entity for the given method.
2039 static ir_entity *create_trampoline(be_main_env_t *be, ir_entity *method)
2041 ir_type *type = get_entity_type(method);
2042 ident *old_id = get_entity_ld_ident(method);
2043 ident *id = id_mangle3("", old_id, "$stub");
2044 ir_type *parent = be->pic_trampolines_type;
2045 ir_entity *ent = new_entity(parent, old_id, type);
2046 set_entity_ld_ident(ent, id);
2047 set_entity_visibility(ent, ir_visibility_private);
2053 * Returns the trampoline entity for the given method.
2055 static ir_entity *get_trampoline(be_main_env_t *env, ir_entity *method)
2057 ir_entity *result = (ir_entity*)pmap_get(env->ent_trampoline_map, method);
2058 if (result == NULL) {
2059 result = create_trampoline(env, method);
2060 pmap_insert(env->ent_trampoline_map, method, result);
2066 static ir_entity *create_pic_symbol(be_main_env_t *be, ir_entity *entity)
2068 ident *old_id = get_entity_ld_ident(entity);
2069 ident *id = id_mangle3("", old_id, "$non_lazy_ptr");
2070 ir_type *e_type = get_entity_type(entity);
2071 ir_type *type = new_type_pointer(e_type);
2072 ir_type *parent = be->pic_symbols_type;
2073 ir_entity *ent = new_entity(parent, old_id, type);
2074 set_entity_ld_ident(ent, id);
2075 set_entity_visibility(ent, ir_visibility_private);
2080 static ir_entity *get_pic_symbol(be_main_env_t *env, ir_entity *entity)
2082 ir_entity *result = (ir_entity*)pmap_get(env->ent_pic_symbol_map, entity);
2083 if (result == NULL) {
2084 result = create_pic_symbol(env, entity);
2085 pmap_insert(env->ent_pic_symbol_map, entity, result);
2094 * Returns non-zero if a given entity can be accessed using a relative address.
2096 static int can_address_relative(ir_entity *entity)
2098 return get_entity_visibility(entity) != ir_visibility_external
2099 && !(get_entity_linkage(entity) & IR_LINKAGE_MERGE);
2102 static ir_node *get_pic_base(ir_graph *irg)
2104 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
2105 if (arch_env->impl->get_pic_base == NULL)
2107 return arch_env->impl->get_pic_base(irg);
2110 /** patches SymConsts to work in position independent code */
2111 static void fix_pic_symconsts(ir_node *node, void *data)
2113 ir_graph *irg = get_irn_irg(node);
2114 be_main_env_t *be = be_get_irg_main_env(irg);
2124 arity = get_irn_arity(node);
2125 for (i = 0; i < arity; ++i) {
2127 ir_node *pred = get_irn_n(node, i);
2129 ir_entity *pic_symbol;
2130 ir_node *pic_symconst;
2132 if (!is_SymConst(pred))
2135 entity = get_SymConst_entity(pred);
2136 block = get_nodes_block(pred);
2138 /* calls can jump to relative addresses, so we can directly jump to
2139 the (relatively) known call address or the trampoline */
2140 if (i == 1 && is_Call(node)) {
2141 ir_entity *trampoline;
2142 ir_node *trampoline_const;
2144 if (can_address_relative(entity))
2147 dbgi = get_irn_dbg_info(pred);
2148 trampoline = get_trampoline(be, entity);
2149 trampoline_const = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
2151 set_irn_n(node, i, trampoline_const);
2155 /* everything else is accessed relative to EIP */
2156 mode = get_irn_mode(pred);
2157 pic_base = get_pic_base(irg);
2159 /* all ok now for locally constructed stuff */
2160 if (can_address_relative(entity)) {
2161 ir_node *add = new_r_Add(block, pic_base, pred, mode);
2163 /* make sure the walker doesn't visit this add again */
2164 mark_irn_visited(add);
2165 set_irn_n(node, i, add);
2169 /* get entry from pic symbol segment */
2170 dbgi = get_irn_dbg_info(pred);
2171 pic_symbol = get_pic_symbol(be, entity);
2172 pic_symconst = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
2174 add = new_r_Add(block, pic_base, pic_symconst, mode);
2175 mark_irn_visited(add);
2177 /* we need an extra indirection for global data outside our current
2178 module. The loads are always safe and can therefore float
2179 and need no memory input */
2180 load = new_r_Load(block, new_r_NoMem(irg), add, mode, cons_floats);
2181 load_res = new_r_Proj(load, mode, pn_Load_res);
2183 set_irn_n(node, i, load_res);
2187 be_abi_irg_t *be_abi_introduce(ir_graph *irg)
2189 be_abi_irg_t *env = XMALLOCZ(be_abi_irg_t);
2190 ir_node *old_frame = get_irg_frame(irg);
2191 be_options_t *options = be_get_irg_options(irg);
2192 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
2193 ir_entity *entity = get_irg_entity(irg);
2194 ir_type *method_type = get_entity_type(entity);
2195 be_irg_t *birg = be_birg_from_irg(irg);
2196 struct obstack *obst = &birg->obst;
2202 /* determine allocatable registers */
2203 assert(birg->allocatable_regs == NULL);
2204 birg->allocatable_regs = rbitset_obstack_alloc(obst, arch_env->n_registers);
2205 for (r = 0; r < arch_env->n_registers; ++r) {
2206 const arch_register_t *reg = &arch_env->registers[r];
2207 if ( !(reg->type & arch_register_type_ignore)) {
2208 rbitset_set(birg->allocatable_regs, r);
2212 /* break here if backend provides a custom API.
2213 * Note: we shouldn't have to setup any be_abi_irg_t* stuff at all,
2214 * but need more cleanup to make this work
2216 be_set_irg_abi(irg, env);
2218 be_omit_fp = options->omit_fp;
2220 env->dce_survivor = new_survive_dce();
2221 env->keep_map = pmap_create();
2222 env->call = be_abi_call_new(arch_env->sp->reg_class);
2223 arch_env_get_call_abi(arch_env, method_type, env->call);
2225 env->init_sp = dummy = new_r_Dummy(irg, arch_env->sp->reg_class->mode);
2226 env->calls = NEW_ARR_F(ir_node*, 0);
2229 irg_walk_graph(irg, fix_pic_symconsts, NULL, env);
2232 /* Lower all call nodes in the IRG. */
2236 Beware: init backend abi call object after processing calls,
2237 otherwise some information might be not yet available.
2239 env->cb = env->call->cb->init(env->call, irg);
2241 /* Process the IRG */
2244 /* fix call inputs for state registers */
2245 fix_call_state_inputs(irg);
2247 /* We don't need the keep map anymore. */
2248 pmap_destroy(env->keep_map);
2249 env->keep_map = NULL;
2251 /* calls array is not needed anymore */
2252 DEL_ARR_F(env->calls);
2255 /* reroute the stack origin of the calls to the true stack origin. */
2256 exchange(dummy, env->init_sp);
2257 exchange(old_frame, get_irg_frame(irg));
2259 /* Make some important node pointers survive the dead node elimination. */
2260 survive_dce_register_irn(env->dce_survivor, &env->init_sp);
2261 foreach_pmap(env->regs, ent) {
2262 survive_dce_register_irn(env->dce_survivor, (ir_node **) &ent->value);
2265 env->call->cb->done(env->cb);
2270 void be_abi_free(ir_graph *irg)
2272 be_abi_irg_t *env = be_get_irg_abi(irg);
2274 if (env->call != NULL)
2275 be_abi_call_free(env->call);
2276 if (env->dce_survivor != NULL)
2277 free_survive_dce(env->dce_survivor);
2278 if (env->regs != NULL)
2279 pmap_destroy(env->regs);
2282 be_set_irg_abi(irg, NULL);
2285 void be_put_allocatable_regs(const ir_graph *irg,
2286 const arch_register_class_t *cls, bitset_t *bs)
2288 be_irg_t *birg = be_birg_from_irg(irg);
2289 unsigned *allocatable_regs = birg->allocatable_regs;
2292 assert(bitset_size(bs) == cls->n_regs);
2293 bitset_clear_all(bs);
2294 for (i = 0; i < cls->n_regs; ++i) {
2295 const arch_register_t *reg = &cls->regs[i];
2296 if (rbitset_is_set(allocatable_regs, reg->global_index))
2301 unsigned be_get_n_allocatable_regs(const ir_graph *irg,
2302 const arch_register_class_t *cls)
2304 bitset_t *bs = bitset_alloca(cls->n_regs);
2305 be_put_allocatable_regs(irg, cls, bs);
2306 return bitset_popcount(bs);
2309 void be_set_allocatable_regs(const ir_graph *irg,
2310 const arch_register_class_t *cls,
2311 unsigned *raw_bitset)
2313 be_irg_t *birg = be_birg_from_irg(irg);
2314 unsigned *allocatable_regs = birg->allocatable_regs;
2317 rbitset_clear_all(raw_bitset, cls->n_regs);
2318 for (i = 0; i < cls->n_regs; ++i) {
2319 const arch_register_t *reg = &cls->regs[i];
2320 if (rbitset_is_set(allocatable_regs, reg->global_index))
2321 rbitset_set(raw_bitset, i);
2325 ir_node *be_abi_get_callee_save_irn(be_abi_irg_t *abi, const arch_register_t *reg)
2327 assert(reg->type & arch_register_type_callee_save);
2328 assert(pmap_contains(abi->regs, (void *) reg));
2329 return (ir_node*)pmap_get(abi->regs, (void *) reg);
2332 ir_node *be_abi_get_ignore_irn(be_abi_irg_t *abi, const arch_register_t *reg)
2334 assert(reg->type & arch_register_type_ignore);
2335 assert(pmap_contains(abi->regs, (void *) reg));
2336 return (ir_node*)pmap_get(abi->regs, (void *) reg);
2339 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_abi);
2340 void be_init_abi(void)
2342 FIRM_DBG_REGISTER(dbg, "firm.be.abi");