2 * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Backend ABI implementation.
23 * @author Sebastian Hack, Michael Beck
32 #include "irgraph_t.h"
35 #include "iredges_t.h"
38 #include "irprintf_t.h"
45 #include "raw_bitset.h"
56 #include "bessaconstr.h"
58 #include "betranshlp.h"
60 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
62 typedef struct be_abi_call_arg_t {
63 unsigned is_res : 1; /**< 1: the call argument is a return value. 0: it's a call parameter. */
64 unsigned in_reg : 1; /**< 1: this argument is transmitted in registers. */
65 unsigned on_stack : 1; /**< 1: this argument is transmitted on the stack. */
66 unsigned callee : 1; /**< 1: someone called us. 0: We call another function */
69 const arch_register_t *reg;
72 unsigned alignment; /**< stack alignment */
73 unsigned space_before; /**< allocate space before */
74 unsigned space_after; /**< allocate space after */
77 struct be_abi_call_t {
78 be_abi_call_flags_t flags; /**< Flags describing the ABI behavior on calls */
79 int pop; /**< number of bytes the stack frame is shrinked by the callee on return. */
80 const be_abi_callbacks_t *cb;
81 ir_type *between_type;
83 const arch_register_class_t *cls_addr; /**< register class of the call address */
87 * The ABI information for the current graph.
90 be_abi_call_t *call; /**< The ABI call information. */
92 ir_node *init_sp; /**< The node representing the stack pointer
93 at the start of the function. */
95 ir_node *start; /**< The be_Start params node. */
96 pmap *regs; /**< A map of all callee-save and ignore regs to
97 their Projs to the RegParams node. */
99 int start_block_bias; /**< The stack bias at the end of the start block. */
101 pmap *keep_map; /**< mapping blocks to keep nodes. */
103 ir_node **calls; /**< flexible array containing all be_Call nodes */
106 static ir_heights_t *ir_heights;
108 /** Flag: if set, try to omit the frame pointer in all routines. */
109 static int be_omit_fp = 1;
112 _ ____ ___ ____ _ _ _ _
113 / \ | __ )_ _| / ___|__ _| | | |__ __ _ ___| | _____
114 / _ \ | _ \| | | | / _` | | | '_ \ / _` |/ __| |/ / __|
115 / ___ \| |_) | | | |__| (_| | | | |_) | (_| | (__| <\__ \
116 /_/ \_\____/___| \____\__,_|_|_|_.__/ \__,_|\___|_|\_\___/
118 These callbacks are used by the backend to set the parameters
119 for a specific call type.
123 * Set compare function: compares two ABI call object arguments.
125 static int cmp_call_arg(const void *a, const void *b, size_t n)
127 const be_abi_call_arg_t *p = (const be_abi_call_arg_t*)a;
128 const be_abi_call_arg_t *q = (const be_abi_call_arg_t*)b;
130 return !(p->is_res == q->is_res && p->pos == q->pos && p->callee == q->callee);
134 * Get an ABI call object argument.
136 * @param call the abi call
137 * @param is_res true for call results, false for call arguments
138 * @param pos position of the argument
139 * @param callee context type - if we are callee or caller
141 static be_abi_call_arg_t *get_call_arg(be_abi_call_t *call, int is_res, int pos, int callee)
143 be_abi_call_arg_t arg;
146 memset(&arg, 0, sizeof(arg));
151 hash = is_res * 128 + pos;
153 return (be_abi_call_arg_t*)set_find(call->params, &arg, sizeof(arg), hash);
157 * Set an ABI call object argument.
159 static void remember_call_arg(be_abi_call_arg_t *arg, be_abi_call_t *call, be_abi_context_t context)
161 unsigned hash = arg->is_res * 128 + arg->pos;
162 if (context & ABI_CONTEXT_CALLEE) {
164 set_insert(call->params, arg, sizeof(*arg), hash);
166 if (context & ABI_CONTEXT_CALLER) {
168 set_insert(call->params, arg, sizeof(*arg), hash);
172 /* Set the flags for a call. */
173 void be_abi_call_set_flags(be_abi_call_t *call, be_abi_call_flags_t flags, const be_abi_callbacks_t *cb)
179 /* Sets the number of bytes the stackframe is shrinked by the callee on return */
180 void be_abi_call_set_pop(be_abi_call_t *call, int pop)
186 /* Set register class for call address */
187 void be_abi_call_set_call_address_reg_class(be_abi_call_t *call, const arch_register_class_t *cls)
189 call->cls_addr = cls;
193 void be_abi_call_param_stack(be_abi_call_t *call, int arg_pos,
194 ir_mode *load_mode, unsigned alignment,
195 unsigned space_before, unsigned space_after,
196 be_abi_context_t context)
198 be_abi_call_arg_t arg;
199 memset(&arg, 0, sizeof(arg));
200 assert(alignment > 0 && "Alignment must be greater than 0");
202 arg.load_mode = load_mode;
203 arg.alignment = alignment;
204 arg.space_before = space_before;
205 arg.space_after = space_after;
209 remember_call_arg(&arg, call, context);
212 void be_abi_call_param_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
214 be_abi_call_arg_t arg;
215 memset(&arg, 0, sizeof(arg));
222 remember_call_arg(&arg, call, context);
225 void be_abi_call_res_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
227 be_abi_call_arg_t arg;
228 memset(&arg, 0, sizeof(arg));
235 remember_call_arg(&arg, call, context);
238 /* Get the flags of a ABI call object. */
239 be_abi_call_flags_t be_abi_call_get_flags(const be_abi_call_t *call)
245 * Constructor for a new ABI call object.
247 * @param cls_addr register class of the call address
249 * @return the new ABI call object
251 static be_abi_call_t *be_abi_call_new(const arch_register_class_t *cls_addr)
253 be_abi_call_t *call = XMALLOCZ(be_abi_call_t);
256 call->params = new_set(cmp_call_arg, 16);
258 call->cls_addr = cls_addr;
260 call->flags.bits.try_omit_fp = be_omit_fp;
266 * Destructor for an ABI call object.
268 static void be_abi_call_free(be_abi_call_t *call)
270 del_set(call->params);
275 * Initializes the frame layout from parts
277 * @param frame the stack layout that will be initialized
278 * @param args the stack argument layout type
279 * @param between the between layout type
280 * @param locals the method frame type
281 * @param param_map an array mapping method argument positions to the stack argument type
283 * @return the initialized stack layout
285 static be_stack_layout_t *stack_frame_init(be_stack_layout_t *frame, ir_type *args,
286 ir_type *between, ir_type *locals,
287 ir_entity *param_map[])
289 frame->arg_type = args;
290 frame->between_type = between;
291 frame->frame_type = locals;
292 frame->initial_offset = 0;
293 frame->initial_bias = 0;
294 frame->order[1] = between;
295 frame->param_map = param_map;
297 /* typical decreasing stack: locals have the
298 * lowest addresses, arguments the highest */
299 frame->order[0] = locals;
300 frame->order[2] = args;
311 Adjustment of the calls inside a graph.
316 * Transform a call node into a be_Call node.
318 * @param env The ABI environment for the current irg.
319 * @param irn The call node.
320 * @param curr_sp The stack pointer node to use.
321 * @return The stack pointer after the call.
323 static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
325 ir_graph *irg = get_irn_irg(irn);
326 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
327 ir_type *call_tp = get_Call_type(irn);
328 ir_node *call_ptr = get_Call_ptr(irn);
329 size_t n_params = get_method_n_params(call_tp);
330 ir_node *curr_mem = get_Call_mem(irn);
331 ir_node *bl = get_nodes_block(irn);
333 const arch_register_t *sp = arch_env->sp;
334 be_abi_call_t *call = be_abi_call_new(sp->reg_class);
335 ir_mode *mach_mode = sp->reg_class->mode;
336 int no_alloc = call->flags.bits.frame_is_setup_on_call;
337 int n_res = get_method_n_ress(call_tp);
338 int do_seq = call->flags.bits.store_args_sequential && !no_alloc;
340 ir_node *res_proj = NULL;
341 int n_reg_params = 0;
342 int n_stack_params = 0;
345 const arch_register_t **states = NEW_ARR_F(const arch_register_t*, 0);
346 const arch_register_t **destroyed_regs = NEW_ARR_F(const arch_register_t*, 0);
350 int n_reg_results = 0;
351 const ir_edge_t *edge;
353 int *stack_param_idx;
354 int i, n, destroy_all_regs;
359 /* Let the isa fill out the abi description for that call node. */
360 arch_env_get_call_abi(arch_env, call_tp, call);
362 /* Insert code to put the stack arguments on the stack. */
363 assert(get_Call_n_params(irn) == n_params);
364 stack_param_idx = ALLOCAN(int, n_params);
365 for (p = 0; p < n_params; ++p) {
366 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
369 int arg_size = get_type_size_bytes(get_method_param_type(call_tp, p));
371 stack_size += round_up2(arg->space_before, arg->alignment);
372 stack_size += round_up2(arg_size, arg->alignment);
373 stack_size += round_up2(arg->space_after, arg->alignment);
375 stack_param_idx[n_stack_params++] = p;
379 /* Collect all arguments which are passed in registers. */
380 reg_param_idxs = ALLOCAN(int, n_params);
381 for (p = 0; p < n_params; ++p) {
382 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
383 if (arg && arg->in_reg) {
384 reg_param_idxs[n_reg_params++] = p;
389 * If the stack is decreasing and we do not want to store sequentially,
390 * or someone else allocated the call frame
391 * we allocate as much space on the stack all parameters need, by
392 * moving the stack pointer along the stack's direction.
394 * Note: we also have to do this for stack_size == 0, because we may have
395 * to adjust stack alignment for the call.
397 if (!do_seq && !no_alloc) {
398 curr_sp = be_new_IncSP(sp, bl, curr_sp, stack_size, 1);
401 dbgi = get_irn_dbg_info(irn);
402 /* If there are some parameters which shall be passed on the stack. */
403 if (n_stack_params > 0) {
405 ir_node **in = ALLOCAN(ir_node*, n_stack_params+1);
409 * Reverse list of stack parameters if call arguments are from left to right.
410 * We must them reverse again if they are pushed (not stored) and the stack
411 * direction is downwards.
413 if (call->flags.bits.left_to_right ^ do_seq) {
414 for (i = 0; i < n_stack_params >> 1; ++i) {
415 int other = n_stack_params - i - 1;
416 int tmp = stack_param_idx[i];
417 stack_param_idx[i] = stack_param_idx[other];
418 stack_param_idx[other] = tmp;
422 curr_mem = get_Call_mem(irn);
424 in[n_in++] = curr_mem;
427 for (i = 0; i < n_stack_params; ++i) {
428 int p = stack_param_idx[i];
429 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
430 ir_node *param = get_Call_param(irn, p);
431 ir_node *addr = curr_sp;
433 ir_type *param_type = get_method_param_type(call_tp, p);
434 int param_size = get_type_size_bytes(param_type) + arg->space_after;
437 * If we wanted to build the arguments sequentially,
438 * the stack pointer for the next must be incremented,
439 * and the memory value propagated.
443 addr = curr_sp = be_new_IncSP(sp, bl, curr_sp,
444 param_size + arg->space_before, 0);
445 add_irn_dep(curr_sp, curr_mem);
447 curr_ofs += arg->space_before;
448 curr_ofs = round_up2(curr_ofs, arg->alignment);
450 /* Make the expression to compute the argument's offset. */
452 ir_mode *constmode = mach_mode;
453 if (mode_is_reference(mach_mode)) {
456 addr = new_r_Const_long(irg, constmode, curr_ofs);
457 addr = new_r_Add(bl, curr_sp, addr, mach_mode);
461 /* Insert a store for primitive arguments. */
462 if (is_atomic_type(param_type)) {
463 ir_node *mem_input = do_seq ? curr_mem : new_r_NoMem(irg);
464 ir_node *store = new_rd_Store(dbgi, bl, mem_input, addr, param, cons_none);
465 mem = new_r_Proj(store, mode_M, pn_Store_M);
467 /* Make a mem copy for compound arguments. */
470 assert(mode_is_reference(get_irn_mode(param)));
471 copy = new_rd_CopyB(dbgi, bl, curr_mem, addr, param, param_type);
472 mem = new_r_Proj(copy, mode_M, pn_CopyB_M);
475 curr_ofs += param_size;
483 /* We need the sync only, if we didn't build the stores sequentially. */
485 if (n_stack_params >= 1) {
486 curr_mem = new_r_Sync(bl, n_in, in);
488 curr_mem = get_Call_mem(irn);
493 /* check for the return_twice property */
494 destroy_all_regs = 0;
495 if (is_SymConst_addr_ent(call_ptr)) {
496 ir_entity *ent = get_SymConst_entity(call_ptr);
498 if (get_entity_additional_properties(ent) & mtp_property_returns_twice)
499 destroy_all_regs = 1;
501 ir_type *call_tp = get_Call_type(irn);
503 if (get_method_additional_properties(call_tp) & mtp_property_returns_twice)
504 destroy_all_regs = 1;
507 /* Put caller save into the destroyed set and state registers in the states
509 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
511 const arch_register_class_t *cls = &arch_env->register_classes[i];
512 for (j = 0; j < cls->n_regs; ++j) {
513 const arch_register_t *reg = arch_register_for_index(cls, j);
515 /* even if destroyed all is specified, neither SP nor FP are
516 * destroyed (else bad things will happen) */
517 if (reg == arch_env->sp || reg == arch_env->bp)
520 if (reg->type & arch_register_type_state) {
521 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
522 ARR_APP1(const arch_register_t*, states, reg);
523 /* we're already in the destroyed set so no need for further
527 if (destroy_all_regs || (reg->type & arch_register_type_caller_save)) {
528 if (!(reg->type & arch_register_type_ignore)) {
529 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
535 /* search the largest result proj number */
536 res_projs = ALLOCANZ(ir_node*, n_res);
538 foreach_out_edge(irn, edge) {
539 const ir_edge_t *res_edge;
540 ir_node *irn = get_edge_src_irn(edge);
542 if (!is_Proj(irn) || get_Proj_proj(irn) != pn_Call_T_result)
545 foreach_out_edge(irn, res_edge) {
547 ir_node *res = get_edge_src_irn(res_edge);
549 assert(is_Proj(res));
551 proj = get_Proj_proj(res);
552 assert(proj < n_res);
553 assert(res_projs[proj] == NULL);
554 res_projs[proj] = res;
560 /** TODO: this is not correct for cases where return values are passed
561 * on the stack, but no known ABI does this currently...
563 n_reg_results = n_res;
566 in = ALLOCAN(ir_node*, n_reg_params + ARR_LEN(states));
568 /* make the back end call node and set its register requirements. */
569 for (i = 0; i < n_reg_params; ++i) {
570 in[n_ins++] = get_Call_param(irn, reg_param_idxs[i]);
573 /* add state registers ins */
574 for (s = 0; s < ARR_LEN(states); ++s) {
575 const arch_register_t *reg = states[s];
576 const arch_register_class_t *cls = arch_register_get_class(reg);
578 ir_node *regnode = be_abi_reg_map_get(env->regs, reg);
579 ir_fprintf(stderr, "Adding %+F\n", regnode);
581 ir_node *regnode = new_r_Unknown(irg, arch_register_class_mode(cls));
582 in[n_ins++] = regnode;
584 assert(n_ins == (int) (n_reg_params + ARR_LEN(states)));
586 /* ins collected, build the call */
587 if (env->call->flags.bits.call_has_imm && is_SymConst(call_ptr)) {
589 low_call = be_new_Call(dbgi, irg, bl, curr_mem, curr_sp, curr_sp,
590 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
591 n_ins, in, get_Call_type(irn));
592 be_Call_set_entity(low_call, get_SymConst_entity(call_ptr));
595 low_call = be_new_Call(dbgi, irg, bl, curr_mem, curr_sp, call_ptr,
596 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
597 n_ins, in, get_Call_type(irn));
599 be_Call_set_pop(low_call, call->pop);
601 /* put the call into the list of all calls for later processing */
602 ARR_APP1(ir_node *, env->calls, low_call);
604 /* create new stack pointer */
605 curr_sp = new_r_Proj(low_call, get_irn_mode(curr_sp), pn_be_Call_sp);
606 be_set_constr_single_reg_out(low_call, pn_be_Call_sp, sp,
607 arch_register_req_type_ignore | arch_register_req_type_produces_sp);
608 arch_set_irn_register(curr_sp, sp);
610 /* now handle results */
611 for (i = 0; i < n_res; ++i) {
613 ir_node *proj = res_projs[i];
614 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
616 /* returns values on stack not supported yet */
620 shift the proj number to the right, since we will drop the
621 unspeakable Proj_T from the Call. Therefore, all real argument
622 Proj numbers must be increased by pn_be_Call_first_res
624 pn = i + pn_be_Call_first_res;
627 ir_type *res_type = get_method_res_type(call_tp, i);
628 ir_mode *mode = get_type_mode(res_type);
629 proj = new_r_Proj(low_call, mode, pn);
632 set_Proj_pred(proj, low_call);
633 set_Proj_proj(proj, pn);
637 /* remove register from destroyed regs */
639 size_t n = ARR_LEN(destroyed_regs);
640 for (j = 0; j < n; ++j) {
641 if (destroyed_regs[j] == arg->reg) {
642 destroyed_regs[j] = destroyed_regs[n-1];
643 ARR_SHRINKLEN(destroyed_regs,n-1);
651 Set the register class of the call address to
652 the backend provided class (default: stack pointer class)
654 be_node_set_reg_class_in(low_call, n_be_Call_ptr, call->cls_addr);
656 DBG((dbg, LEVEL_3, "\tcreated backend call %+F\n", low_call));
658 /* Set the register classes and constraints of the Call parameters. */
659 for (i = 0; i < n_reg_params; ++i) {
660 int index = reg_param_idxs[i];
661 be_abi_call_arg_t *arg = get_call_arg(call, 0, index, 0);
662 assert(arg->reg != NULL);
664 be_set_constr_single_reg_in(low_call, n_be_Call_first_arg + i,
665 arg->reg, arch_register_req_type_none);
668 /* Set the register constraints of the results. */
669 for (i = 0; i < n_res; ++i) {
670 ir_node *proj = res_projs[i];
671 const be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
672 int pn = get_Proj_proj(proj);
675 be_set_constr_single_reg_out(low_call, pn, arg->reg,
676 arch_register_req_type_none);
677 arch_set_irn_register(proj, arg->reg);
679 exchange(irn, low_call);
681 /* kill the ProjT node */
682 if (res_proj != NULL) {
686 /* Make additional projs for the caller save registers
687 and the Keep node which keeps them alive. */
693 int curr_res_proj = pn_be_Call_first_res + n_reg_results;
696 n_ins = ARR_LEN(destroyed_regs) + n_reg_results + 1;
697 in = ALLOCAN(ir_node *, n_ins);
699 /* also keep the stack pointer */
700 set_irn_link(curr_sp, (void*) sp);
703 for (d = 0; d < ARR_LEN(destroyed_regs); ++d) {
704 const arch_register_t *reg = destroyed_regs[d];
705 ir_node *proj = new_r_Proj(low_call, reg->reg_class->mode, curr_res_proj);
707 /* memorize the register in the link field. we need afterwards to set the register class of the keep correctly. */
708 be_set_constr_single_reg_out(low_call, curr_res_proj, reg,
709 arch_register_req_type_none);
710 arch_set_irn_register(proj, reg);
712 set_irn_link(proj, (void*) reg);
717 for (i = 0; i < n_reg_results; ++i) {
718 ir_node *proj = res_projs[i];
719 const arch_register_t *reg = arch_get_irn_register(proj);
720 set_irn_link(proj, (void*) reg);
725 /* create the Keep for the caller save registers */
726 keep = be_new_Keep(bl, n, in);
727 for (i = 0; i < n; ++i) {
728 const arch_register_t *reg = (const arch_register_t*)get_irn_link(in[i]);
729 be_node_set_reg_class_in(keep, i, reg->reg_class);
733 /* Clean up the stack. */
734 assert(stack_size >= call->pop);
735 stack_size -= call->pop;
737 if (stack_size > 0) {
738 ir_node *mem_proj = NULL;
740 foreach_out_edge(low_call, edge) {
741 ir_node *irn = get_edge_src_irn(edge);
742 if (is_Proj(irn) && get_Proj_proj(irn) == pn_Call_M) {
749 mem_proj = new_r_Proj(low_call, mode_M, pn_be_Call_M_regular);
750 keep_alive(mem_proj);
753 /* Clean up the stack frame or revert alignment fixes if we allocated it */
755 curr_sp = be_new_IncSP(sp, bl, curr_sp, -stack_size, 0);
758 be_abi_call_free(call);
761 DEL_ARR_F(destroyed_regs);
767 * Adjust the size of a node representing a stack alloc or free for the minimum stack alignment.
769 * @param alignment the minimum stack alignment
770 * @param size the node containing the non-aligned size
771 * @param block the block where new nodes are allocated on
772 * @param dbg debug info for new nodes
774 * @return a node representing the aligned size
776 static ir_node *adjust_alloc_size(unsigned stack_alignment, ir_node *size,
777 ir_node *block, dbg_info *dbg)
779 if (stack_alignment > 1) {
785 assert(is_po2(stack_alignment));
787 mode = get_irn_mode(size);
788 tv = new_tarval_from_long(stack_alignment-1, mode);
789 irg = get_Block_irg(block);
790 mask = new_r_Const(irg, tv);
791 size = new_rd_Add(dbg, block, size, mask, mode);
793 tv = new_tarval_from_long(-(long)stack_alignment, mode);
794 mask = new_r_Const(irg, tv);
795 size = new_rd_And(dbg, block, size, mask, mode);
801 * The alloca is transformed into a back end alloca node and connected to the stack nodes.
803 static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp)
805 ir_node *block = get_nodes_block(alloc);
806 ir_graph *irg = get_Block_irg(block);
807 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
808 ir_node *alloc_mem = NULL;
809 ir_node *alloc_res = NULL;
810 ir_type *type = get_Alloc_type(alloc);
813 const ir_edge_t *edge;
818 unsigned stack_alignment;
820 /* all non-stack Alloc nodes should already be lowered before the backend */
821 assert(get_Alloc_where(alloc) == stack_alloc);
823 foreach_out_edge(alloc, edge) {
824 ir_node *irn = get_edge_src_irn(edge);
826 assert(is_Proj(irn));
827 switch (get_Proj_proj(irn)) {
839 /* Beware: currently Alloc nodes without a result might happen,
840 only escape analysis kills them and this phase runs only for object
841 oriented source. We kill the Alloc here. */
842 if (alloc_res == NULL && alloc_mem) {
843 exchange(alloc_mem, get_Alloc_mem(alloc));
847 dbg = get_irn_dbg_info(alloc);
848 count = get_Alloc_count(alloc);
850 /* we might need to multiply the count with the element size */
851 if (type != firm_unknown_type && get_type_size_bytes(type) != 1) {
852 ir_mode *mode = get_irn_mode(count);
853 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type),
855 ir_node *cnst = new_rd_Const(dbg, irg, tv);
856 size = new_rd_Mul(dbg, block, count, cnst, mode);
861 /* The stack pointer will be modified in an unknown manner.
862 We cannot omit it. */
863 env->call->flags.bits.try_omit_fp = 0;
865 stack_alignment = 1 << arch_env->stack_alignment;
866 size = adjust_alloc_size(stack_alignment, size, block, dbg);
867 new_alloc = be_new_AddSP(arch_env->sp, block, curr_sp, size);
868 set_irn_dbg_info(new_alloc, dbg);
870 if (alloc_mem != NULL) {
874 addsp_mem = new_r_Proj(new_alloc, mode_M, pn_be_AddSP_M);
876 /* We need to sync the output mem of the AddSP with the input mem
877 edge into the alloc node. */
878 ins[0] = get_Alloc_mem(alloc);
880 sync = new_r_Sync(block, 2, ins);
882 exchange(alloc_mem, sync);
885 exchange(alloc, new_alloc);
887 /* fix projnum of alloca res */
888 set_Proj_proj(alloc_res, pn_be_AddSP_res);
890 curr_sp = new_r_Proj(new_alloc, get_irn_mode(curr_sp), pn_be_AddSP_sp);
897 * The Free is transformed into a back end free node and connected to the stack nodes.
899 static ir_node *adjust_free(be_abi_irg_t *env, ir_node *free, ir_node *curr_sp)
901 ir_node *block = get_nodes_block(free);
902 ir_graph *irg = get_irn_irg(free);
903 ir_type *type = get_Free_type(free);
904 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
905 ir_mode *sp_mode = arch_env->sp->reg_class->mode;
906 dbg_info *dbg = get_irn_dbg_info(free);
907 ir_node *subsp, *mem, *res, *size, *sync;
909 unsigned stack_alignment;
911 /* all non-stack-alloc Free nodes should already be lowered before the
913 assert(get_Free_where(free) == stack_alloc);
915 /* we might need to multiply the size with the element size */
916 if (type != firm_unknown_type && get_type_size_bytes(type) != 1) {
917 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type), mode_Iu);
918 ir_node *cnst = new_rd_Const(dbg, irg, tv);
919 ir_node *mul = new_rd_Mul(dbg, block, get_Free_size(free),
923 size = get_Free_size(free);
926 stack_alignment = 1 << arch_env->stack_alignment;
927 size = adjust_alloc_size(stack_alignment, size, block, dbg);
929 /* The stack pointer will be modified in an unknown manner.
930 We cannot omit it. */
931 env->call->flags.bits.try_omit_fp = 0;
932 subsp = be_new_SubSP(arch_env->sp, block, curr_sp, size);
933 set_irn_dbg_info(subsp, dbg);
935 mem = new_r_Proj(subsp, mode_M, pn_be_SubSP_M);
936 res = new_r_Proj(subsp, sp_mode, pn_be_SubSP_sp);
938 /* we need to sync the memory */
939 in[0] = get_Free_mem(free);
941 sync = new_r_Sync(block, 2, in);
943 /* and make the AddSP dependent on the former memory */
944 add_irn_dep(subsp, get_Free_mem(free));
947 exchange(free, sync);
954 * Check if a node is somehow data dependent on another one.
955 * both nodes must be in the same basic block.
956 * @param n1 The first node.
957 * @param n2 The second node.
958 * @return 1, if n1 is data dependent (transitively) on n2, 0 if not.
960 static int dependent_on(ir_node *n1, ir_node *n2)
962 assert(get_nodes_block(n1) == get_nodes_block(n2));
964 return heights_reachable_in_block(ir_heights, n1, n2);
967 static int cmp_call_dependency(const void *c1, const void *c2)
969 ir_node *n1 = *(ir_node **) c1;
970 ir_node *n2 = *(ir_node **) c2;
974 Classical qsort() comparison function behavior:
975 0 if both elements are equal
976 1 if second is "smaller" that first
977 -1 if first is "smaller" that second
979 if (dependent_on(n1, n2))
982 if (dependent_on(n2, n1))
985 /* The nodes have no depth order, but we need a total order because qsort()
988 * Additionally, we need to respect transitive dependencies. Consider a
989 * Call a depending on Call b and an independent Call c.
990 * We MUST NOT order c > a and b > c. */
991 h1 = get_irn_height(ir_heights, n1);
992 h2 = get_irn_height(ir_heights, n2);
993 if (h1 < h2) return -1;
994 if (h1 > h2) return 1;
995 /* Same height, so use a random (but stable) order */
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 int dir = env->call->flags.bits.left_to_right ? 1 : -1;
1141 int n = get_method_n_params(method_type);
1142 int curr = inc > 0 ? 0 : n - 1;
1143 struct obstack *obst = be_get_be_obst(irg);
1149 ident *id = get_entity_ident(get_irg_entity(irg));
1152 *param_map = map = OALLOCN(obst, ir_entity*, n);
1153 res = new_type_struct(id_mangle_u(id, new_id_from_chars("arg_type", 8)));
1154 for (i = 0; i < n; ++i, curr += inc) {
1155 ir_type *param_type = get_method_param_type(method_type, curr);
1156 be_abi_call_arg_t *arg = get_call_arg(call, 0, curr, 1);
1159 if (arg->on_stack) {
1160 if (val_param_tp != NULL) {
1161 /* the entity was already created, create a copy in the param type */
1162 ir_entity *val_ent = get_method_value_param_ent(method_type, i);
1163 arg->stack_ent = copy_entity_own(val_ent, res);
1164 set_entity_link(val_ent, arg->stack_ent);
1165 set_entity_link(arg->stack_ent, NULL);
1167 /* create a new entity */
1168 snprintf(buf, sizeof(buf), "param_%d", i);
1169 arg->stack_ent = new_entity(res, new_id_from_str(buf), param_type);
1171 ofs += arg->space_before;
1172 ofs = round_up2(ofs, arg->alignment);
1173 set_entity_offset(arg->stack_ent, ofs);
1174 ofs += arg->space_after;
1175 ofs += get_type_size_bytes(param_type);
1176 map[i] = arg->stack_ent;
1179 set_type_size_bytes(res, ofs);
1180 set_type_state(res, layout_fixed);
1185 const arch_register_t *reg;
1189 static int cmp_regs(const void *a, const void *b)
1191 const reg_node_map_t *p = (const reg_node_map_t*)a;
1192 const reg_node_map_t *q = (const reg_node_map_t*)b;
1194 if (p->reg->reg_class == q->reg->reg_class)
1195 return p->reg->index - q->reg->index;
1197 return p->reg->reg_class < q->reg->reg_class ? -1 : +1;
1200 static void reg_map_to_arr(reg_node_map_t *res, pmap *reg_map)
1203 size_t n = pmap_count(reg_map);
1206 foreach_pmap(reg_map, ent) {
1207 res[i].reg = (const arch_register_t*)ent->key;
1208 res[i].irn = (ir_node*)ent->value;
1212 qsort(res, n, sizeof(res[0]), cmp_regs);
1216 * Creates a be_Return for a Return node.
1218 * @param @env the abi environment
1219 * @param irn the Return node or NULL if there was none
1220 * @param bl the block where the be_Retun should be placed
1221 * @param mem the current memory
1222 * @param n_res number of return results
1224 static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl,
1225 ir_node *mem, int n_res)
1227 be_abi_call_t *call = env->call;
1228 ir_graph *irg = get_Block_irg(bl);
1229 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1231 pmap *reg_map = pmap_create();
1232 ir_node *keep = (ir_node*)pmap_get(env->keep_map, bl);
1239 const arch_register_t **regs;
1243 get the valid stack node in this block.
1244 If we had a call in that block there is a Keep constructed by process_calls()
1245 which points to the last stack modification in that block. we'll use
1246 it then. Else we use the stack from the start block and let
1247 the ssa construction fix the usage.
1249 stack = be_abi_reg_map_get(env->regs, arch_env->sp);
1251 stack = get_irn_n(keep, 0);
1253 remove_End_keepalive(get_irg_end(irg), keep);
1256 /* Insert results for Return into the register map. */
1257 for (i = 0; i < n_res; ++i) {
1258 ir_node *res = get_Return_res(irn, i);
1259 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1260 assert(arg->in_reg && "return value must be passed in register");
1261 pmap_insert(reg_map, (void *) arg->reg, res);
1264 /* Add uses of the callee save registers. */
1265 foreach_pmap(env->regs, ent) {
1266 const arch_register_t *reg = (const arch_register_t*)ent->key;
1267 if (reg->type & (arch_register_type_callee_save | arch_register_type_ignore))
1268 pmap_insert(reg_map, ent->key, ent->value);
1271 be_abi_reg_map_set(reg_map, arch_env->sp, stack);
1274 Maximum size of the in array for Return nodes is
1275 return args + callee save/ignore registers + memory + stack pointer
1277 in_max = pmap_count(reg_map) + n_res + 2;
1279 in = ALLOCAN(ir_node*, in_max);
1280 regs = ALLOCAN(arch_register_t const*, in_max);
1283 in[1] = be_abi_reg_map_get(reg_map, arch_env->sp);
1285 regs[1] = arch_env->sp;
1288 /* clear SP entry, since it has already been grown. */
1289 pmap_insert(reg_map, (void *) arch_env->sp, NULL);
1290 for (i = 0; i < n_res; ++i) {
1291 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1293 in[n] = be_abi_reg_map_get(reg_map, arg->reg);
1294 regs[n++] = arg->reg;
1296 /* Clear the map entry to mark the register as processed. */
1297 be_abi_reg_map_set(reg_map, arg->reg, NULL);
1300 /* grow the rest of the stuff. */
1301 foreach_pmap(reg_map, ent) {
1303 in[n] = (ir_node*)ent->value;
1304 regs[n++] = (const arch_register_t*)ent->key;
1308 /* The in array for the new back end return is now ready. */
1310 dbgi = get_irn_dbg_info(irn);
1314 /* we have to pop the shadow parameter in in case of struct returns */
1316 ret = be_new_Return(dbgi, irg, bl, n_res, pop, n, in);
1318 /* Set the register classes of the return's parameter accordingly. */
1319 for (i = 0; i < n; ++i) {
1320 if (regs[i] == NULL)
1323 be_set_constr_single_reg_in(ret, i, regs[i], arch_register_req_type_none);
1326 /* Free the space of the Epilog's in array and the register <-> proj map. */
1327 pmap_destroy(reg_map);
1332 typedef struct ent_pos_pair ent_pos_pair;
1333 struct ent_pos_pair {
1334 ir_entity *ent; /**< a value param entity */
1335 int pos; /**< its parameter number */
1336 ent_pos_pair *next; /**< for linking */
1339 typedef struct lower_frame_sels_env_t {
1340 ent_pos_pair *value_param_list; /**< the list of all value param entities */
1341 ir_node *frame; /**< the current frame */
1342 const arch_register_class_t *sp_class; /**< register class of the stack pointer */
1343 const arch_register_class_t *link_class; /**< register class of the link pointer */
1344 ir_type *value_tp; /**< the value type if any */
1345 ir_type *frame_tp; /**< the frame type */
1346 int static_link_pos; /**< argument number of the hidden static link */
1347 } lower_frame_sels_env_t;
1350 * Return an entity from the backend for an value param entity.
1352 * @param ent an value param type entity
1353 * @param ctx context
1355 static ir_entity *get_argument_entity(ir_entity *ent, lower_frame_sels_env_t *ctx)
1357 ir_entity *argument_ent = (ir_entity*)get_entity_link(ent);
1359 if (argument_ent == NULL) {
1360 /* we have NO argument entity yet: This is bad, as we will
1361 * need one for backing store.
1364 ir_type *frame_tp = ctx->frame_tp;
1365 unsigned offset = get_type_size_bytes(frame_tp);
1366 ir_type *tp = get_entity_type(ent);
1367 unsigned align = get_type_alignment_bytes(tp);
1369 offset += align - 1;
1370 offset &= ~(align - 1);
1372 argument_ent = copy_entity_own(ent, frame_tp);
1374 /* must be automatic to set a fixed layout */
1375 set_entity_offset(argument_ent, offset);
1376 offset += get_type_size_bytes(tp);
1378 set_type_size_bytes(frame_tp, offset);
1379 set_entity_link(ent, argument_ent);
1381 return argument_ent;
1384 * Walker: Replaces Sels of frame type and
1385 * value param type entities by FrameAddress.
1386 * Links all used entities.
1388 static void lower_frame_sels_walker(ir_node *irn, void *data)
1390 lower_frame_sels_env_t *ctx = (lower_frame_sels_env_t*)data;
1393 ir_node *ptr = get_Sel_ptr(irn);
1395 if (ptr == ctx->frame) {
1396 ir_entity *ent = get_Sel_entity(irn);
1397 ir_node *bl = get_nodes_block(irn);
1400 int is_value_param = 0;
1402 if (get_entity_owner(ent) == ctx->value_tp) {
1405 /* replace by its copy from the argument type */
1406 pos = get_struct_member_index(ctx->value_tp, ent);
1407 ent = get_argument_entity(ent, ctx);
1410 nw = be_new_FrameAddr(ctx->sp_class, bl, ctx->frame, ent);
1413 /* check, if it's a param Sel and if have not seen this entity before */
1414 if (is_value_param && get_entity_link(ent) == NULL) {
1420 ARR_APP1(ent_pos_pair, ctx->value_param_list, pair);
1422 set_entity_link(ent, ctx->value_param_list);
1429 * Check if a value parameter is transmitted as a register.
1430 * This might happen if the address of an parameter is taken which is
1431 * transmitted in registers.
1433 * Note that on some architectures this case must be handled specially
1434 * because the place of the backing store is determined by their ABI.
1436 * In the default case we move the entity to the frame type and create
1437 * a backing store into the first block.
1439 static void fix_address_of_parameter_access(be_abi_irg_t *env, ir_graph *irg,
1440 ent_pos_pair *value_param_list)
1442 be_abi_call_t *call = env->call;
1443 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1444 ent_pos_pair *entry, *new_list;
1446 int i, n = ARR_LEN(value_param_list);
1449 for (i = 0; i < n; ++i) {
1450 int pos = value_param_list[i].pos;
1451 be_abi_call_arg_t *arg = get_call_arg(call, 0, pos, 1);
1454 DBG((dbg, LEVEL_2, "\targ #%d need backing store\n", pos));
1455 value_param_list[i].next = new_list;
1456 new_list = &value_param_list[i];
1459 if (new_list != NULL) {
1460 /* ok, change the graph */
1461 ir_node *start_bl = get_irg_start_block(irg);
1462 ir_node *first_bl = get_first_block_succ(start_bl);
1463 ir_node *frame, *imem, *nmem, *store, *mem, *args;
1464 optimization_state_t state;
1467 assert(first_bl && first_bl != start_bl);
1468 /* we had already removed critical edges, so the following
1469 assertion should be always true. */
1470 assert(get_Block_n_cfgpreds(first_bl) == 1);
1472 /* now create backing stores */
1473 frame = get_irg_frame(irg);
1474 imem = get_irg_initial_mem(irg);
1476 save_optimization_state(&state);
1478 nmem = new_r_Proj(get_irg_start(irg), mode_M, pn_Start_M);
1479 restore_optimization_state(&state);
1481 /* reroute all edges to the new memory source */
1482 edges_reroute(imem, nmem);
1486 args = get_irg_args(irg);
1487 for (entry = new_list; entry != NULL; entry = entry->next) {
1489 ir_type *tp = get_entity_type(entry->ent);
1490 ir_mode *mode = get_type_mode(tp);
1493 /* address for the backing store */
1494 addr = be_new_FrameAddr(arch_env->sp->reg_class, first_bl, frame, entry->ent);
1497 mem = new_r_Proj(store, mode_M, pn_Store_M);
1499 /* the backing store itself */
1500 store = new_r_Store(first_bl, mem, addr,
1501 new_r_Proj(args, mode, i), cons_none);
1503 /* the new memory Proj gets the last Proj from store */
1504 set_Proj_pred(nmem, store);
1505 set_Proj_proj(nmem, pn_Store_M);
1507 /* move all entities to the frame type */
1508 frame_tp = get_irg_frame_type(irg);
1509 offset = get_type_size_bytes(frame_tp);
1511 /* we will add new entities: set the layout to undefined */
1512 assert(get_type_state(frame_tp) == layout_fixed);
1513 set_type_state(frame_tp, layout_undefined);
1514 for (entry = new_list; entry != NULL; entry = entry->next) {
1515 ir_entity *ent = entry->ent;
1517 /* If the entity is still on the argument type, move it to the
1519 * This happens if the value_param type was build due to compound
1521 if (get_entity_owner(ent) != frame_tp) {
1522 ir_type *tp = get_entity_type(ent);
1523 unsigned align = get_type_alignment_bytes(tp);
1525 offset += align - 1;
1526 offset &= ~(align - 1);
1527 set_entity_owner(ent, frame_tp);
1528 /* must be automatic to set a fixed layout */
1529 set_entity_offset(ent, offset);
1530 offset += get_type_size_bytes(tp);
1533 set_type_size_bytes(frame_tp, offset);
1534 /* fix the layout again */
1535 set_type_state(frame_tp, layout_fixed);
1540 * The start block has no jump, instead it has an initial exec Proj.
1541 * The backend wants to handle all blocks the same way, so we replace
1542 * the out cfg edge with a real jump.
1544 static void fix_start_block(ir_graph *irg)
1546 ir_node *initial_X = get_irg_initial_exec(irg);
1547 ir_node *start_block = get_irg_start_block(irg);
1548 ir_node *jmp = new_r_Jmp(start_block);
1550 assert(is_Proj(initial_X));
1551 exchange(initial_X, jmp);
1552 set_irg_initial_exec(irg, new_r_Bad(irg));
1556 * Update the entity of Sels to the outer value parameters.
1558 static void update_outer_frame_sels(ir_node *irn, void *env)
1560 lower_frame_sels_env_t *ctx = (lower_frame_sels_env_t*)env;
1567 ptr = get_Sel_ptr(irn);
1568 if (! is_arg_Proj(ptr))
1570 if (get_Proj_proj(ptr) != ctx->static_link_pos)
1572 ent = get_Sel_entity(irn);
1574 if (get_entity_owner(ent) == ctx->value_tp) {
1575 /* replace by its copy from the argument type */
1576 pos = get_struct_member_index(ctx->value_tp, ent);
1577 ent = get_argument_entity(ent, ctx);
1578 set_Sel_entity(irn, ent);
1580 /* check, if we have not seen this entity before */
1581 if (get_entity_link(ent) == NULL) {
1587 ARR_APP1(ent_pos_pair, ctx->value_param_list, pair);
1589 set_entity_link(ent, ctx->value_param_list);
1595 * Fix access to outer local variables.
1597 static void fix_outer_variable_access(be_abi_irg_t *env,
1598 lower_frame_sels_env_t *ctx)
1604 for (i = get_class_n_members(ctx->frame_tp) - 1; i >= 0; --i) {
1605 ir_entity *ent = get_class_member(ctx->frame_tp, i);
1607 if (! is_method_entity(ent))
1610 irg = get_entity_irg(ent);
1615 * FIXME: find the number of the static link parameter
1616 * for now we assume 0 here
1618 ctx->static_link_pos = 0;
1620 irg_walk_graph(irg, NULL, update_outer_frame_sels, ctx);
1625 * Modify the irg itself and the frame type.
1627 static void modify_irg(ir_graph *irg)
1629 be_abi_irg_t *env = be_get_irg_abi(irg);
1630 be_abi_call_t *call = env->call;
1631 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1632 const arch_register_t *sp = arch_env->sp;
1633 ir_type *method_type = get_entity_type(get_irg_entity(irg));
1634 be_irg_t *birg = be_birg_from_irg(irg);
1635 struct obstack *obst = be_get_be_obst(irg);
1636 be_stack_layout_t *stack_layout = be_get_irg_stack_layout(irg);
1639 ir_node *new_mem_proj;
1645 unsigned frame_size;
1648 const arch_register_t *fp_reg;
1649 ir_node *frame_pointer;
1653 const ir_edge_t *edge;
1654 ir_type *arg_type, *bet_type, *tp;
1655 lower_frame_sels_env_t ctx;
1656 ir_entity **param_map;
1658 DBG((dbg, LEVEL_1, "introducing abi on %+F\n", irg));
1660 old_mem = get_irg_initial_mem(irg);
1662 irp_reserve_resources(irp, IR_RESOURCE_ENTITY_LINK);
1664 /* set the links of all frame entities to NULL, we use it
1665 to detect if an entity is already linked in the value_param_list */
1666 tp = get_method_value_param_type(method_type);
1669 /* clear the links of the clone type, let the
1670 original entities point to its clones */
1671 for (i = get_struct_n_members(tp) - 1; i >= 0; --i) {
1672 ir_entity *mem = get_struct_member(tp, i);
1673 set_entity_link(mem, NULL);
1677 arg_type = compute_arg_type(env, irg, call, method_type, tp, ¶m_map);
1679 /* Convert the Sel nodes in the irg to frame addr nodes: */
1680 ctx.value_param_list = NEW_ARR_F(ent_pos_pair, 0);
1681 ctx.frame = get_irg_frame(irg);
1682 ctx.sp_class = arch_env->sp->reg_class;
1683 ctx.link_class = arch_env->link_class;
1684 ctx.frame_tp = get_irg_frame_type(irg);
1686 /* layout the stackframe now */
1687 if (get_type_state(ctx.frame_tp) == layout_undefined) {
1688 default_layout_compound_type(ctx.frame_tp);
1691 /* we will possible add new entities to the frame: set the layout to undefined */
1692 assert(get_type_state(ctx.frame_tp) == layout_fixed);
1693 set_type_state(ctx.frame_tp, layout_undefined);
1695 irg_walk_graph(irg, lower_frame_sels_walker, NULL, &ctx);
1697 /* fix the frame type layout again */
1698 set_type_state(ctx.frame_tp, layout_fixed);
1699 /* align stackframe to 4 byte */
1700 frame_size = get_type_size_bytes(ctx.frame_tp);
1701 if (frame_size % 4 != 0) {
1702 set_type_size_bytes(ctx.frame_tp, frame_size + 4 - (frame_size % 4));
1705 env->regs = pmap_create();
1707 n_params = get_method_n_params(method_type);
1708 args = OALLOCNZ(obst, ir_node*, n_params);
1711 * for inner function we must now fix access to outer frame entities.
1713 fix_outer_variable_access(env, &ctx);
1715 /* Check if a value parameter is transmitted as a register.
1716 * This might happen if the address of an parameter is taken which is
1717 * transmitted in registers.
1719 * Note that on some architectures this case must be handled specially
1720 * because the place of the backing store is determined by their ABI.
1722 * In the default case we move the entity to the frame type and create
1723 * a backing store into the first block.
1725 fix_address_of_parameter_access(env, irg, ctx.value_param_list);
1727 DEL_ARR_F(ctx.value_param_list);
1728 irp_free_resources(irp, IR_RESOURCE_ENTITY_LINK);
1730 /* Fill the argument vector */
1731 arg_tuple = get_irg_args(irg);
1732 foreach_out_edge(arg_tuple, edge) {
1733 ir_node *irn = get_edge_src_irn(edge);
1734 if (! is_Anchor(irn)) {
1735 int nr = get_Proj_proj(irn);
1737 DBG((dbg, LEVEL_2, "\treading arg: %d -> %+F\n", nr, irn));
1741 stack_layout->sp_relative = call->flags.bits.try_omit_fp;
1742 bet_type = call->cb->get_between_type(irg);
1743 stack_frame_init(stack_layout, arg_type, bet_type,
1744 get_irg_frame_type(irg), param_map);
1746 /* Count the register params and add them to the number of Projs for the RegParams node */
1747 for (i = 0; i < n_params; ++i) {
1748 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1749 if (arg->in_reg && args[i]) {
1750 assert(arg->reg != sp && "cannot use stack pointer as parameter register");
1751 assert(i == get_Proj_proj(args[i]));
1753 /* For now, associate the register with the old Proj from Start representing that argument. */
1754 pmap_insert(env->regs, (void *) arg->reg, args[i]);
1755 DBG((dbg, LEVEL_2, "\targ #%d -> reg %s\n", i, arg->reg->name));
1759 /* Collect all callee-save registers */
1760 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
1761 const arch_register_class_t *cls = &arch_env->register_classes[i];
1762 for (j = 0; j < cls->n_regs; ++j) {
1763 const arch_register_t *reg = &cls->regs[j];
1764 if (reg->type & (arch_register_type_callee_save | arch_register_type_state)) {
1765 pmap_insert(env->regs, (void *) reg, NULL);
1770 fp_reg = call->flags.bits.try_omit_fp ? arch_env->sp : arch_env->bp;
1771 rbitset_clear(birg->allocatable_regs, fp_reg->global_index);
1773 /* handle start block here (place a jump in the block) */
1774 fix_start_block(irg);
1776 pmap_insert(env->regs, (void *) sp, NULL);
1777 pmap_insert(env->regs, (void *) arch_env->bp, NULL);
1778 start_bl = get_irg_start_block(irg);
1779 env->start = be_new_Start(NULL, start_bl, pmap_count(env->regs) + 1);
1780 set_irg_start(irg, env->start);
1783 * make proj nodes for the callee save registers.
1784 * memorize them, since Return nodes get those as inputs.
1786 * Note, that if a register corresponds to an argument, the regs map
1787 * contains the old Proj from start for that argument.
1789 rm = ALLOCAN(reg_node_map_t, pmap_count(env->regs));
1790 reg_map_to_arr(rm, env->regs);
1791 for (i = 0, n = pmap_count(env->regs); i < n; ++i) {
1792 const arch_register_t *reg = rm[i].reg;
1793 ir_mode *mode = reg->reg_class->mode;
1795 arch_register_req_type_t add_type = arch_register_req_type_none;
1799 add_type |= arch_register_req_type_produces_sp;
1800 if (!rbitset_is_set(birg->allocatable_regs, reg->global_index)) {
1801 add_type |= arch_register_req_type_ignore;
1805 proj = new_r_Proj(env->start, mode, nr + 1);
1806 pmap_insert(env->regs, (void *) reg, proj);
1807 be_set_constr_single_reg_out(env->start, nr + 1, reg, add_type);
1808 arch_set_irn_register(proj, reg);
1810 DBG((dbg, LEVEL_2, "\tregister save proj #%d -> reg %s\n", nr, reg->name));
1813 /* create a new initial memory proj */
1814 assert(is_Proj(old_mem));
1815 arch_set_out_register_req(env->start, 0, arch_no_register_req);
1816 new_mem_proj = new_r_Proj(env->start, mode_M, 0);
1818 set_irg_initial_mem(irg, mem);
1820 env->init_sp = be_abi_reg_map_get(env->regs, sp);
1822 /* set new frame_pointer */
1823 frame_pointer = be_abi_reg_map_get(env->regs, fp_reg);
1824 set_irg_frame(irg, frame_pointer);
1826 /* rewire old mem users to new mem */
1827 exchange(old_mem, mem);
1829 /* keep the mem (for functions with an endless loop = no return) */
1832 set_irg_initial_mem(irg, mem);
1834 /* Now, introduce stack param nodes for all parameters passed on the stack */
1835 for (i = 0; i < n_params; ++i) {
1836 ir_node *arg_proj = args[i];
1837 ir_node *repl = NULL;
1839 if (arg_proj != NULL) {
1840 be_abi_call_arg_t *arg;
1841 ir_type *param_type;
1842 int nr = get_Proj_proj(arg_proj);
1845 nr = MIN(nr, n_params);
1846 arg = get_call_arg(call, 0, nr, 1);
1847 param_type = get_method_param_type(method_type, nr);
1850 repl = (ir_node*)pmap_get(env->regs, arg->reg);
1851 } else if (arg->on_stack) {
1852 ir_node *addr = be_new_FrameAddr(sp->reg_class, start_bl, frame_pointer, arg->stack_ent);
1854 /* For atomic parameters which are actually used, we create a Load node. */
1855 if (is_atomic_type(param_type) && get_irn_n_edges(args[i]) > 0) {
1856 ir_mode *mode = get_type_mode(param_type);
1857 ir_mode *load_mode = arg->load_mode;
1859 ir_node *load = new_r_Load(start_bl, new_r_NoMem(irg), addr, load_mode, cons_floats);
1860 repl = new_r_Proj(load, load_mode, pn_Load_res);
1862 if (mode != load_mode) {
1863 repl = new_r_Conv(start_bl, repl, mode);
1866 /* The stack parameter is not primitive (it is a struct or array),
1867 * we thus will create a node representing the parameter's address
1873 assert(repl != NULL);
1875 /* Beware: the mode of the register parameters is always the mode of the register class
1876 which may be wrong. Add Conv's then. */
1877 mode = get_irn_mode(args[i]);
1878 if (mode != get_irn_mode(repl)) {
1879 repl = new_r_Conv(get_nodes_block(repl), repl, mode);
1881 exchange(args[i], repl);
1885 /* the arg proj is not needed anymore now and should be only used by the anchor */
1886 assert(get_irn_n_edges(arg_tuple) == 1);
1887 kill_node(arg_tuple);
1888 set_irg_args(irg, new_r_Bad(irg));
1890 /* All Return nodes hang on the End node, so look for them there. */
1891 end = get_irg_end_block(irg);
1892 for (i = 0, n = get_Block_n_cfgpreds(end); i < n; ++i) {
1893 ir_node *irn = get_Block_cfgpred(end, i);
1895 if (is_Return(irn)) {
1896 ir_node *blk = get_nodes_block(irn);
1897 ir_node *mem = get_Return_mem(irn);
1898 ir_node *ret = create_be_return(env, irn, blk, mem, get_Return_n_ress(irn));
1903 /* if we have endless loops here, n might be <= 0. Do NOT create a be_Return then,
1904 the code is dead and will never be executed. */
1907 /** Fix the state inputs of calls that still hang on unknowns */
1908 static void fix_call_state_inputs(ir_graph *irg)
1910 be_abi_irg_t *env = be_get_irg_abi(irg);
1911 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1913 arch_register_t **stateregs = NEW_ARR_F(arch_register_t*, 0);
1915 /* Collect caller save registers */
1916 n = arch_env->n_register_classes;
1917 for (i = 0; i < n; ++i) {
1919 const arch_register_class_t *cls = &arch_env->register_classes[i];
1920 for (j = 0; j < cls->n_regs; ++j) {
1921 const arch_register_t *reg = arch_register_for_index(cls, j);
1922 if (reg->type & arch_register_type_state) {
1923 ARR_APP1(arch_register_t*, stateregs, (arch_register_t *)reg);
1928 n = ARR_LEN(env->calls);
1929 n_states = ARR_LEN(stateregs);
1930 for (i = 0; i < n; ++i) {
1932 ir_node *call = env->calls[i];
1934 arity = get_irn_arity(call);
1936 /* the state reg inputs are the last n inputs of the calls */
1937 for (s = 0; s < n_states; ++s) {
1938 int inp = arity - n_states + s;
1939 const arch_register_t *reg = stateregs[s];
1940 ir_node *regnode = be_abi_reg_map_get(env->regs, reg);
1942 set_irn_n(call, inp, regnode);
1946 DEL_ARR_F(stateregs);
1950 * Create a trampoline entity for the given method.
1952 static ir_entity *create_trampoline(be_main_env_t *be, ir_entity *method)
1954 ir_type *type = get_entity_type(method);
1955 ident *old_id = get_entity_ld_ident(method);
1956 ident *id = id_mangle3("", old_id, "$stub");
1957 ir_type *parent = be->pic_trampolines_type;
1958 ir_entity *ent = new_entity(parent, old_id, type);
1959 set_entity_ld_ident(ent, id);
1960 set_entity_visibility(ent, ir_visibility_private);
1966 * Returns the trampoline entity for the given method.
1968 static ir_entity *get_trampoline(be_main_env_t *env, ir_entity *method)
1970 ir_entity *result = (ir_entity*)pmap_get(env->ent_trampoline_map, method);
1971 if (result == NULL) {
1972 result = create_trampoline(env, method);
1973 pmap_insert(env->ent_trampoline_map, method, result);
1979 static ir_entity *create_pic_symbol(be_main_env_t *be, ir_entity *entity)
1981 ident *old_id = get_entity_ld_ident(entity);
1982 ident *id = id_mangle3("", old_id, "$non_lazy_ptr");
1983 ir_type *e_type = get_entity_type(entity);
1984 ir_type *type = new_type_pointer(e_type);
1985 ir_type *parent = be->pic_symbols_type;
1986 ir_entity *ent = new_entity(parent, old_id, type);
1987 set_entity_ld_ident(ent, id);
1988 set_entity_visibility(ent, ir_visibility_private);
1993 static ir_entity *get_pic_symbol(be_main_env_t *env, ir_entity *entity)
1995 ir_entity *result = (ir_entity*)pmap_get(env->ent_pic_symbol_map, entity);
1996 if (result == NULL) {
1997 result = create_pic_symbol(env, entity);
1998 pmap_insert(env->ent_pic_symbol_map, entity, result);
2007 * Returns non-zero if a given entity can be accessed using a relative address.
2009 static int can_address_relative(ir_entity *entity)
2011 return get_entity_visibility(entity) != ir_visibility_external
2012 && !(get_entity_linkage(entity) & IR_LINKAGE_MERGE);
2015 static ir_node *get_pic_base(ir_graph *irg)
2017 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
2018 if (arch_env->impl->get_pic_base == NULL)
2020 return arch_env->impl->get_pic_base(irg);
2023 /** patches SymConsts to work in position independent code */
2024 static void fix_pic_symconsts(ir_node *node, void *data)
2026 ir_graph *irg = get_irn_irg(node);
2027 be_main_env_t *be = be_get_irg_main_env(irg);
2037 arity = get_irn_arity(node);
2038 for (i = 0; i < arity; ++i) {
2040 ir_node *pred = get_irn_n(node, i);
2042 ir_entity *pic_symbol;
2043 ir_node *pic_symconst;
2045 if (!is_SymConst(pred))
2048 entity = get_SymConst_entity(pred);
2049 block = get_nodes_block(pred);
2051 /* calls can jump to relative addresses, so we can directly jump to
2052 the (relatively) known call address or the trampoline */
2053 if (i == 1 && is_Call(node)) {
2054 ir_entity *trampoline;
2055 ir_node *trampoline_const;
2057 if (can_address_relative(entity))
2060 dbgi = get_irn_dbg_info(pred);
2061 trampoline = get_trampoline(be, entity);
2062 trampoline_const = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
2064 set_irn_n(node, i, trampoline_const);
2068 /* everything else is accessed relative to EIP */
2069 mode = get_irn_mode(pred);
2070 pic_base = get_pic_base(irg);
2072 /* all ok now for locally constructed stuff */
2073 if (can_address_relative(entity)) {
2074 ir_node *add = new_r_Add(block, pic_base, pred, mode);
2076 /* make sure the walker doesn't visit this add again */
2077 mark_irn_visited(add);
2078 set_irn_n(node, i, add);
2082 /* get entry from pic symbol segment */
2083 dbgi = get_irn_dbg_info(pred);
2084 pic_symbol = get_pic_symbol(be, entity);
2085 pic_symconst = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
2087 add = new_r_Add(block, pic_base, pic_symconst, mode);
2088 mark_irn_visited(add);
2090 /* we need an extra indirection for global data outside our current
2091 module. The loads are always safe and can therefore float
2092 and need no memory input */
2093 load = new_r_Load(block, new_r_NoMem(irg), add, mode, cons_floats);
2094 load_res = new_r_Proj(load, mode, pn_Load_res);
2096 set_irn_n(node, i, load_res);
2100 be_abi_irg_t *be_abi_introduce(ir_graph *irg)
2102 be_abi_irg_t *env = XMALLOCZ(be_abi_irg_t);
2103 ir_node *old_frame = get_irg_frame(irg);
2104 be_options_t *options = be_get_irg_options(irg);
2105 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
2106 ir_entity *entity = get_irg_entity(irg);
2107 ir_type *method_type = get_entity_type(entity);
2108 be_irg_t *birg = be_birg_from_irg(irg);
2109 struct obstack *obst = &birg->obst;
2110 ir_node *dummy = new_r_Dummy(irg,
2111 arch_env->sp->reg_class->mode);
2114 /* determine allocatable registers */
2115 assert(birg->allocatable_regs == NULL);
2116 birg->allocatable_regs = rbitset_obstack_alloc(obst, arch_env->n_registers);
2117 for (r = 0; r < arch_env->n_registers; ++r) {
2118 const arch_register_t *reg = &arch_env->registers[r];
2119 if ( !(reg->type & arch_register_type_ignore)) {
2120 rbitset_set(birg->allocatable_regs, r);
2124 /* break here if backend provides a custom API.
2125 * Note: we shouldn't have to setup any be_abi_irg_t* stuff at all,
2126 * but need more cleanup to make this work
2128 be_set_irg_abi(irg, env);
2130 be_omit_fp = options->omit_fp;
2132 env->keep_map = pmap_create();
2133 env->call = be_abi_call_new(arch_env->sp->reg_class);
2134 arch_env_get_call_abi(arch_env, method_type, env->call);
2136 env->init_sp = dummy;
2137 env->calls = NEW_ARR_F(ir_node*, 0);
2140 irg_walk_graph(irg, fix_pic_symconsts, NULL, env);
2143 /* Lower all call nodes in the IRG. */
2146 /* Process the IRG */
2149 /* fix call inputs for state registers */
2150 fix_call_state_inputs(irg);
2152 /* We don't need the keep map anymore. */
2153 pmap_destroy(env->keep_map);
2154 env->keep_map = NULL;
2156 /* calls array is not needed anymore */
2157 DEL_ARR_F(env->calls);
2160 /* reroute the stack origin of the calls to the true stack origin. */
2161 exchange(dummy, env->init_sp);
2162 exchange(old_frame, get_irg_frame(irg));
2167 void be_abi_free(ir_graph *irg)
2169 be_abi_irg_t *env = be_get_irg_abi(irg);
2171 if (env->call != NULL)
2172 be_abi_call_free(env->call);
2173 if (env->regs != NULL)
2174 pmap_destroy(env->regs);
2177 be_set_irg_abi(irg, NULL);
2181 * called after nodes have been transformed so some node references can be
2182 * replaced with new nodes
2184 void be_abi_transform_fixup(ir_graph *irg)
2186 be_abi_irg_t *abi = be_get_irg_abi(irg);
2189 if (abi == NULL || abi->regs == NULL)
2192 new_regs = pmap_create();
2193 foreach_pmap(abi->regs, entry) {
2194 ir_node *value = (ir_node*)entry->value;
2195 ir_node *transformed = be_transform_node(value);
2196 pmap_insert(new_regs, entry->key, transformed);
2198 pmap_destroy(abi->regs);
2199 abi->regs = new_regs;
2202 void be_put_allocatable_regs(const ir_graph *irg,
2203 const arch_register_class_t *cls, bitset_t *bs)
2205 be_irg_t *birg = be_birg_from_irg(irg);
2206 unsigned *allocatable_regs = birg->allocatable_regs;
2209 assert(bitset_size(bs) == cls->n_regs);
2210 bitset_clear_all(bs);
2211 for (i = 0; i < cls->n_regs; ++i) {
2212 const arch_register_t *reg = &cls->regs[i];
2213 if (rbitset_is_set(allocatable_regs, reg->global_index))
2218 unsigned be_get_n_allocatable_regs(const ir_graph *irg,
2219 const arch_register_class_t *cls)
2221 bitset_t *bs = bitset_alloca(cls->n_regs);
2222 be_put_allocatable_regs(irg, cls, bs);
2223 return bitset_popcount(bs);
2226 void be_set_allocatable_regs(const ir_graph *irg,
2227 const arch_register_class_t *cls,
2228 unsigned *raw_bitset)
2230 be_irg_t *birg = be_birg_from_irg(irg);
2231 unsigned *allocatable_regs = birg->allocatable_regs;
2234 rbitset_clear_all(raw_bitset, cls->n_regs);
2235 for (i = 0; i < cls->n_regs; ++i) {
2236 const arch_register_t *reg = &cls->regs[i];
2237 if (rbitset_is_set(allocatable_regs, reg->global_index))
2238 rbitset_set(raw_bitset, i);
2242 ir_node *be_abi_get_callee_save_irn(ir_graph *irg, const arch_register_t *reg)
2244 const be_abi_irg_t *abi = be_get_irg_abi(irg);
2245 assert(reg->type & arch_register_type_callee_save);
2246 assert(pmap_contains(abi->regs, (void *) reg));
2247 return (ir_node*)pmap_get(abi->regs, (void *) reg);
2250 ir_node *be_abi_get_ignore_irn(ir_graph *irg, const arch_register_t *reg)
2252 const be_abi_irg_t *abi = be_get_irg_abi(irg);
2253 assert(pmap_contains(abi->regs, (void *) reg));
2254 return (ir_node*)pmap_get(abi->regs, (void *) reg);
2257 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_abi);
2258 void be_init_abi(void)
2260 FIRM_DBG_REGISTER(dbg, "firm.be.abi");