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
31 #include "irgraph_t.h"
34 #include "iredges_t.h"
37 #include "irprintf_t.h"
44 #include "raw_bitset.h"
55 #include "bessaconstr.h"
57 #include "betranshlp.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 be_abi_call_t *call; /**< The ABI call information. */
91 ir_node *init_sp; /**< The node representing the stack pointer
92 at the start of the function. */
94 ir_node *start; /**< The be_Start params node. */
95 pmap *regs; /**< A map of all callee-save and ignore regs to
96 their Projs to the RegParams node. */
97 int start_block_bias; /**< The stack bias at the end of the start block. */
99 pmap *keep_map; /**< mapping blocks to keep nodes. */
101 ir_node **calls; /**< flexible array containing all be_Call nodes */
104 static ir_heights_t *ir_heights;
106 /** Flag: if set, try to omit the frame pointer in all routines. */
107 static int be_omit_fp = 1;
109 static ir_node *be_abi_reg_map_get(pmap *map, const arch_register_t *reg)
111 return (ir_node*)pmap_get(map, reg);
114 static void be_abi_reg_map_set(pmap *map, const arch_register_t* reg,
117 pmap_insert(map, reg, node);
121 * Check if the given register is callee save, ie. will be saved by the callee.
123 static bool arch_register_is_callee_save(
124 const arch_env_t *arch_env,
125 const arch_register_t *reg)
127 if (arch_env->impl->register_saved_by)
128 return arch_env->impl->register_saved_by(reg, /*callee=*/1);
133 * Check if the given register is caller save, ie. must be saved by the caller.
135 static bool arch_register_is_caller_save(
136 const arch_env_t *arch_env,
137 const arch_register_t *reg)
139 if (arch_env->impl->register_saved_by)
140 return arch_env->impl->register_saved_by(reg, /*callee=*/0);
147 _ ____ ___ ____ _ _ _ _
148 / \ | __ )_ _| / ___|__ _| | | |__ __ _ ___| | _____
149 / _ \ | _ \| | | | / _` | | | '_ \ / _` |/ __| |/ / __|
150 / ___ \| |_) | | | |__| (_| | | | |_) | (_| | (__| <\__ \
151 /_/ \_\____/___| \____\__,_|_|_|_.__/ \__,_|\___|_|\_\___/
153 These callbacks are used by the backend to set the parameters
154 for a specific call type.
158 * Set compare function: compares two ABI call object arguments.
160 static int cmp_call_arg(const void *a, const void *b, size_t n)
162 const be_abi_call_arg_t *p = (const be_abi_call_arg_t*)a;
163 const be_abi_call_arg_t *q = (const be_abi_call_arg_t*)b;
165 return !(p->is_res == q->is_res && p->pos == q->pos && p->callee == q->callee);
169 * Get an ABI call object argument.
171 * @param call the abi call
172 * @param is_res true for call results, false for call arguments
173 * @param pos position of the argument
174 * @param callee context type - if we are callee or caller
176 static be_abi_call_arg_t *get_call_arg(be_abi_call_t *call, int is_res, int pos, int callee)
178 be_abi_call_arg_t arg;
181 memset(&arg, 0, sizeof(arg));
186 hash = is_res * 128 + pos;
188 return (be_abi_call_arg_t*)set_find(call->params, &arg, sizeof(arg), hash);
192 * Set an ABI call object argument.
194 static void remember_call_arg(be_abi_call_arg_t *arg, be_abi_call_t *call, be_abi_context_t context)
196 unsigned hash = arg->is_res * 128 + arg->pos;
197 if (context & ABI_CONTEXT_CALLEE) {
199 set_insert(call->params, arg, sizeof(*arg), hash);
201 if (context & ABI_CONTEXT_CALLER) {
203 set_insert(call->params, arg, sizeof(*arg), hash);
207 /* Set the flags for a call. */
208 void be_abi_call_set_flags(be_abi_call_t *call, be_abi_call_flags_t flags, const be_abi_callbacks_t *cb)
214 /* Sets the number of bytes the stackframe is shrinked by the callee on return */
215 void be_abi_call_set_pop(be_abi_call_t *call, int pop)
221 /* Set register class for call address */
222 void be_abi_call_set_call_address_reg_class(be_abi_call_t *call, const arch_register_class_t *cls)
224 call->cls_addr = cls;
228 void be_abi_call_param_stack(be_abi_call_t *call, int arg_pos,
229 ir_mode *load_mode, unsigned alignment,
230 unsigned space_before, unsigned space_after,
231 be_abi_context_t context)
233 be_abi_call_arg_t arg;
234 memset(&arg, 0, sizeof(arg));
235 assert(alignment > 0 && "Alignment must be greater than 0");
237 arg.load_mode = load_mode;
238 arg.alignment = alignment;
239 arg.space_before = space_before;
240 arg.space_after = space_after;
244 remember_call_arg(&arg, call, context);
247 void be_abi_call_param_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
249 be_abi_call_arg_t arg;
250 memset(&arg, 0, sizeof(arg));
257 remember_call_arg(&arg, call, context);
260 void be_abi_call_res_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
262 be_abi_call_arg_t arg;
263 memset(&arg, 0, sizeof(arg));
270 remember_call_arg(&arg, call, context);
273 /* Get the flags of a ABI call object. */
274 be_abi_call_flags_t be_abi_call_get_flags(const be_abi_call_t *call)
280 * Constructor for a new ABI call object.
282 * @param cls_addr register class of the call address
284 * @return the new ABI call object
286 static be_abi_call_t *be_abi_call_new(const arch_register_class_t *cls_addr)
288 be_abi_call_t *call = XMALLOCZ(be_abi_call_t);
291 call->params = new_set(cmp_call_arg, 16);
293 call->cls_addr = cls_addr;
295 call->flags.bits.try_omit_fp = be_omit_fp;
301 * Destructor for an ABI call object.
303 static void be_abi_call_free(be_abi_call_t *call)
305 del_set(call->params);
310 * Initializes the frame layout from parts
312 * @param frame the stack layout that will be initialized
313 * @param args the stack argument layout type
314 * @param between the between layout type
315 * @param locals the method frame type
316 * @param param_map an array mapping method argument positions to the stack argument type
318 * @return the initialized stack layout
320 static be_stack_layout_t *stack_frame_init(be_stack_layout_t *frame, ir_type *args,
321 ir_type *between, ir_type *locals,
322 ir_entity *param_map[])
324 frame->arg_type = args;
325 frame->between_type = between;
326 frame->frame_type = locals;
327 frame->initial_offset = 0;
328 frame->initial_bias = 0;
329 frame->order[1] = between;
330 frame->param_map = param_map;
332 /* typical decreasing stack: locals have the
333 * lowest addresses, arguments the highest */
334 frame->order[0] = locals;
335 frame->order[2] = args;
346 Adjustment of the calls inside a graph.
351 * Transform a call node into a be_Call node.
353 * @param env The ABI environment for the current irg.
354 * @param irn The call node.
355 * @param curr_sp The stack pointer node to use.
356 * @return The stack pointer after the call.
358 static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
360 ir_graph *irg = get_irn_irg(irn);
361 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
362 ir_type *call_tp = get_Call_type(irn);
363 ir_node *call_ptr = get_Call_ptr(irn);
364 size_t n_params = get_method_n_params(call_tp);
365 ir_node *curr_mem = get_Call_mem(irn);
366 ir_node *bl = get_nodes_block(irn);
368 const arch_register_t *sp = arch_env->sp;
369 be_abi_call_t *call = be_abi_call_new(sp->reg_class);
370 ir_mode *mach_mode = sp->reg_class->mode;
371 int n_res = get_method_n_ress(call_tp);
373 ir_node *res_proj = NULL;
374 int n_reg_params = 0;
375 int n_stack_params = 0;
378 const arch_register_t **states = NEW_ARR_F(const arch_register_t*, 0);
379 const arch_register_t **destroyed_regs = NEW_ARR_F(const arch_register_t*, 0);
383 int n_reg_results = 0;
384 const ir_edge_t *edge;
386 int *stack_param_idx;
388 int throws_exception;
393 /* Let the isa fill out the abi description for that call node. */
394 arch_env_get_call_abi(arch_env, call_tp, call);
396 /* Insert code to put the stack arguments on the stack. */
397 assert(get_Call_n_params(irn) == n_params);
398 stack_param_idx = ALLOCAN(int, n_params);
399 for (p = 0; p < n_params; ++p) {
400 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
403 int arg_size = get_type_size_bytes(get_method_param_type(call_tp, p));
405 stack_size += round_up2(arg->space_before, arg->alignment);
406 stack_size += round_up2(arg_size, arg->alignment);
407 stack_size += round_up2(arg->space_after, arg->alignment);
409 stack_param_idx[n_stack_params++] = p;
413 /* Collect all arguments which are passed in registers. */
414 reg_param_idxs = ALLOCAN(int, n_params);
415 for (p = 0; p < n_params; ++p) {
416 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
417 if (arg && arg->in_reg) {
418 reg_param_idxs[n_reg_params++] = p;
423 * If the stack is decreasing and we do not want to store sequentially,
424 * or someone else allocated the call frame
425 * we allocate as much space on the stack all parameters need, by
426 * moving the stack pointer along the stack's direction.
428 * Note: we also have to do this for stack_size == 0, because we may have
429 * to adjust stack alignment for the call.
431 curr_sp = be_new_IncSP(sp, bl, curr_sp, stack_size, 1);
433 dbgi = get_irn_dbg_info(irn);
434 /* If there are some parameters which shall be passed on the stack. */
435 if (n_stack_params > 0) {
437 ir_node **in = ALLOCAN(ir_node*, n_stack_params+1);
440 curr_mem = get_Call_mem(irn);
441 in[n_in++] = curr_mem;
443 for (i = 0; i < n_stack_params; ++i) {
444 int p = stack_param_idx[i];
445 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
446 ir_node *param = get_Call_param(irn, p);
447 ir_node *addr = curr_sp;
449 ir_type *param_type = get_method_param_type(call_tp, p);
450 int param_size = get_type_size_bytes(param_type) + arg->space_after;
453 * If we wanted to build the arguments sequentially,
454 * the stack pointer for the next must be incremented,
455 * and the memory value propagated.
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);
470 /* Insert a store for primitive arguments. */
471 if (is_atomic_type(param_type)) {
472 ir_node *nomem = get_irg_no_mem(irg);
473 ir_node *mem_input = nomem;
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;
490 /* We need the sync only, if we didn't build the stores sequentially. */
491 if (n_stack_params >= 1) {
492 curr_mem = new_r_Sync(bl, n_in, in);
494 curr_mem = get_Call_mem(irn);
498 /* Put caller save into the destroyed set and state registers in the states
500 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
502 const arch_register_class_t *cls = &arch_env->register_classes[i];
503 for (j = 0; j < cls->n_regs; ++j) {
504 const arch_register_t *reg = arch_register_for_index(cls, j);
506 /* even if destroyed all is specified, neither SP nor FP are
507 * destroyed (else bad things will happen) */
508 if (reg == arch_env->sp || reg == arch_env->bp)
511 if (reg->type & arch_register_type_state) {
512 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
513 ARR_APP1(const arch_register_t*, states, reg);
514 /* we're already in the destroyed set so no need for further
518 if (arch_register_is_caller_save(arch_env, reg)) {
519 if (!(reg->type & arch_register_type_ignore)) {
520 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
526 /* search the largest result proj number */
527 res_projs = ALLOCANZ(ir_node*, n_res);
529 foreach_out_edge(irn, edge) {
530 const ir_edge_t *res_edge;
531 ir_node *irn = get_edge_src_irn(edge);
533 if (!is_Proj(irn) || get_Proj_proj(irn) != pn_Call_T_result)
536 foreach_out_edge(irn, res_edge) {
538 ir_node *res = get_edge_src_irn(res_edge);
540 assert(is_Proj(res));
542 proj = get_Proj_proj(res);
543 assert(proj < n_res);
544 assert(res_projs[proj] == NULL);
545 res_projs[proj] = res;
551 /** TODO: this is not correct for cases where return values are passed
552 * on the stack, but no known ABI does this currently...
554 n_reg_results = n_res;
557 in = ALLOCAN(ir_node*, n_reg_params + ARR_LEN(states));
559 /* make the back end call node and set its register requirements. */
560 for (i = 0; i < n_reg_params; ++i) {
561 in[n_ins++] = get_Call_param(irn, reg_param_idxs[i]);
564 /* add state registers ins */
565 for (s = 0; s < ARR_LEN(states); ++s) {
566 const arch_register_t *reg = states[s];
567 const arch_register_class_t *cls = reg->reg_class;
568 ir_node *regnode = new_r_Unknown(irg, cls->mode);
569 in[n_ins++] = regnode;
571 assert(n_ins == (int) (n_reg_params + ARR_LEN(states)));
573 /* ins collected, build the call */
574 throws_exception = ir_throws_exception(irn);
575 if (env->call->flags.bits.call_has_imm && is_SymConst(call_ptr)) {
577 low_call = be_new_Call(dbgi, irg, bl, curr_mem, curr_sp, curr_sp,
578 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
579 n_ins, in, get_Call_type(irn));
580 be_Call_set_entity(low_call, get_SymConst_entity(call_ptr));
583 low_call = be_new_Call(dbgi, irg, bl, curr_mem, curr_sp, call_ptr,
584 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
585 n_ins, in, get_Call_type(irn));
587 ir_set_throws_exception(low_call, throws_exception);
588 be_Call_set_pop(low_call, call->pop);
590 /* put the call into the list of all calls for later processing */
591 ARR_APP1(ir_node *, env->calls, low_call);
593 /* create new stack pointer */
594 curr_sp = new_r_Proj(low_call, get_irn_mode(curr_sp), pn_be_Call_sp);
595 be_set_constr_single_reg_out(low_call, pn_be_Call_sp, sp,
596 arch_register_req_type_ignore | arch_register_req_type_produces_sp);
597 arch_set_irn_register(curr_sp, sp);
599 /* now handle results */
600 for (i = 0; i < n_res; ++i) {
601 ir_node *proj = res_projs[i];
602 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
603 long pn = i + pn_be_Call_first_res;
605 /* returns values on stack not supported yet */
609 shift the proj number to the right, since we will drop the
610 unspeakable Proj_T from the Call. Therefore, all real argument
611 Proj numbers must be increased by pn_be_Call_first_res
613 pn = i + pn_be_Call_first_res;
616 ir_type *res_type = get_method_res_type(call_tp, i);
617 ir_mode *mode = get_type_mode(res_type);
618 proj = new_r_Proj(low_call, mode, pn);
621 set_Proj_pred(proj, low_call);
622 set_Proj_proj(proj, pn);
626 /* remove register from destroyed regs */
628 size_t n = ARR_LEN(destroyed_regs);
629 for (j = 0; j < n; ++j) {
630 if (destroyed_regs[j] == arg->reg) {
631 destroyed_regs[j] = destroyed_regs[n-1];
632 ARR_SHRINKLEN(destroyed_regs,n-1);
640 Set the register class of the call address to
641 the backend provided class (default: stack pointer class)
643 be_node_set_reg_class_in(low_call, n_be_Call_ptr, call->cls_addr);
645 DBG((dbg, LEVEL_3, "\tcreated backend call %+F\n", low_call));
647 /* Set the register classes and constraints of the Call parameters. */
648 for (i = 0; i < n_reg_params; ++i) {
649 int index = reg_param_idxs[i];
650 be_abi_call_arg_t *arg = get_call_arg(call, 0, index, 0);
651 assert(arg->reg != NULL);
653 be_set_constr_single_reg_in(low_call, n_be_Call_first_arg + i,
654 arg->reg, arch_register_req_type_none);
657 /* Set the register constraints of the results. */
658 for (i = 0; i < n_res; ++i) {
659 ir_node *proj = res_projs[i];
660 const be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
661 int pn = get_Proj_proj(proj);
664 be_set_constr_single_reg_out(low_call, pn, arg->reg,
665 arch_register_req_type_none);
666 arch_set_irn_register(proj, arg->reg);
668 exchange(irn, low_call);
670 /* kill the ProjT node */
671 if (res_proj != NULL) {
675 /* Make additional projs for the caller save registers
676 and the Keep node which keeps them alive. */
682 int curr_res_proj = pn_be_Call_first_res + n_reg_results;
685 n_ins = ARR_LEN(destroyed_regs) + n_reg_results + 1;
686 in = ALLOCAN(ir_node *, n_ins);
688 /* also keep the stack pointer */
689 set_irn_link(curr_sp, (void*) sp);
692 for (d = 0; d < ARR_LEN(destroyed_regs); ++d) {
693 const arch_register_t *reg = destroyed_regs[d];
694 ir_node *proj = new_r_Proj(low_call, reg->reg_class->mode, curr_res_proj);
696 /* memorize the register in the link field. we need afterwards to set the register class of the keep correctly. */
697 be_set_constr_single_reg_out(low_call, curr_res_proj, reg,
698 arch_register_req_type_none);
699 arch_set_irn_register(proj, reg);
701 set_irn_link(proj, (void*) reg);
706 for (i = 0; i < n_reg_results; ++i) {
707 ir_node *proj = res_projs[i];
708 const arch_register_t *reg = arch_get_irn_register(proj);
709 set_irn_link(proj, (void*) reg);
714 /* create the Keep for the caller save registers */
715 keep = be_new_Keep(bl, n, in);
716 for (i = 0; i < n; ++i) {
717 const arch_register_t *reg = (const arch_register_t*)get_irn_link(in[i]);
718 be_node_set_reg_class_in(keep, i, arch_register_get_class(reg));
722 /* Clean up the stack. */
723 assert(stack_size >= call->pop);
724 stack_size -= call->pop;
726 if (stack_size > 0) {
727 ir_node *mem_proj = NULL;
729 foreach_out_edge(low_call, edge) {
730 ir_node *irn = get_edge_src_irn(edge);
731 if (is_Proj(irn) && get_Proj_proj(irn) == pn_Call_M) {
738 mem_proj = new_r_Proj(low_call, mode_M, pn_be_Call_M);
739 keep_alive(mem_proj);
742 /* Clean up the stack frame or revert alignment fixes if we allocated it */
743 curr_sp = be_new_IncSP(sp, bl, curr_sp, -stack_size, 0);
745 be_abi_call_free(call);
748 DEL_ARR_F(destroyed_regs);
754 * Adjust the size of a node representing a stack alloc or free for the minimum stack alignment.
756 * @param alignment the minimum stack alignment
757 * @param size the node containing the non-aligned size
758 * @param block the block where new nodes are allocated on
759 * @param dbg debug info for new nodes
761 * @return a node representing the aligned size
763 static ir_node *adjust_alloc_size(unsigned stack_alignment, ir_node *size,
764 ir_node *block, dbg_info *dbg)
766 if (stack_alignment > 1) {
772 assert(is_po2(stack_alignment));
774 mode = get_irn_mode(size);
775 tv = new_tarval_from_long(stack_alignment-1, mode);
776 irg = get_Block_irg(block);
777 mask = new_r_Const(irg, tv);
778 size = new_rd_Add(dbg, block, size, mask, mode);
780 tv = new_tarval_from_long(-(long)stack_alignment, mode);
781 mask = new_r_Const(irg, tv);
782 size = new_rd_And(dbg, block, size, mask, mode);
788 * The alloca is transformed into a back end alloca node and connected to the stack nodes.
790 static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp)
792 ir_node *block = get_nodes_block(alloc);
793 ir_graph *irg = get_Block_irg(block);
794 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
795 ir_node *alloc_mem = NULL;
796 ir_node *alloc_res = NULL;
797 ir_type *type = get_Alloc_type(alloc);
800 const ir_edge_t *edge;
805 unsigned stack_alignment;
807 /* all non-stack Alloc nodes should already be lowered before the backend */
808 assert(get_Alloc_where(alloc) == stack_alloc);
810 foreach_out_edge(alloc, edge) {
811 ir_node *irn = get_edge_src_irn(edge);
813 assert(is_Proj(irn));
814 switch (get_Proj_proj(irn)) {
826 /* Beware: currently Alloc nodes without a result might happen,
827 only escape analysis kills them and this phase runs only for object
828 oriented source. We kill the Alloc here. */
829 if (alloc_res == NULL && alloc_mem) {
830 exchange(alloc_mem, get_Alloc_mem(alloc));
834 dbg = get_irn_dbg_info(alloc);
835 count = get_Alloc_count(alloc);
837 /* we might need to multiply the count with the element size */
838 if (!is_unknown_type(type) && get_type_size_bytes(type) != 1) {
839 ir_mode *mode = get_irn_mode(count);
840 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type),
842 ir_node *cnst = new_rd_Const(dbg, irg, tv);
843 size = new_rd_Mul(dbg, block, count, cnst, mode);
848 /* The stack pointer will be modified in an unknown manner.
849 We cannot omit it. */
850 env->call->flags.bits.try_omit_fp = 0;
852 stack_alignment = 1 << arch_env->stack_alignment;
853 size = adjust_alloc_size(stack_alignment, size, block, dbg);
854 new_alloc = be_new_AddSP(arch_env->sp, block, curr_sp, size);
855 set_irn_dbg_info(new_alloc, dbg);
857 if (alloc_mem != NULL) {
861 addsp_mem = new_r_Proj(new_alloc, mode_M, pn_be_AddSP_M);
863 /* We need to sync the output mem of the AddSP with the input mem
864 edge into the alloc node. */
865 ins[0] = get_Alloc_mem(alloc);
867 sync = new_r_Sync(block, 2, ins);
869 exchange(alloc_mem, sync);
872 exchange(alloc, new_alloc);
874 /* fix projnum of alloca res */
875 set_Proj_proj(alloc_res, pn_be_AddSP_res);
877 curr_sp = new_r_Proj(new_alloc, get_irn_mode(curr_sp), pn_be_AddSP_sp);
884 * The Free is transformed into a back end free node and connected to the stack nodes.
886 static ir_node *adjust_free(be_abi_irg_t *env, ir_node *free, ir_node *curr_sp)
888 ir_node *block = get_nodes_block(free);
889 ir_graph *irg = get_irn_irg(free);
890 ir_type *type = get_Free_type(free);
891 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
892 ir_mode *sp_mode = arch_env->sp->reg_class->mode;
893 dbg_info *dbg = get_irn_dbg_info(free);
894 ir_node *subsp, *mem, *res, *size, *sync;
896 unsigned stack_alignment;
898 /* all non-stack-alloc Free nodes should already be lowered before the
900 assert(get_Free_where(free) == stack_alloc);
902 /* we might need to multiply the size with the element size */
903 if (!is_unknown_type(type) && get_type_size_bytes(type) != 1) {
904 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type), mode_Iu);
905 ir_node *cnst = new_rd_Const(dbg, irg, tv);
906 ir_node *mul = new_rd_Mul(dbg, block, get_Free_count(free),
910 size = get_Free_count(free);
913 stack_alignment = 1 << arch_env->stack_alignment;
914 size = adjust_alloc_size(stack_alignment, size, block, dbg);
916 /* The stack pointer will be modified in an unknown manner.
917 We cannot omit it. */
918 env->call->flags.bits.try_omit_fp = 0;
919 subsp = be_new_SubSP(arch_env->sp, block, curr_sp, size);
920 set_irn_dbg_info(subsp, dbg);
922 mem = new_r_Proj(subsp, mode_M, pn_be_SubSP_M);
923 res = new_r_Proj(subsp, sp_mode, pn_be_SubSP_sp);
925 /* we need to sync the memory */
926 in[0] = get_Free_mem(free);
928 sync = new_r_Sync(block, 2, in);
930 /* and make the AddSP dependent on the former memory */
931 add_irn_dep(subsp, get_Free_mem(free));
934 exchange(free, sync);
941 * Check if a node is somehow data dependent on another one.
942 * both nodes must be in the same basic block.
943 * @param n1 The first node.
944 * @param n2 The second node.
945 * @return 1, if n1 is data dependent (transitively) on n2, 0 if not.
947 static int dependent_on(ir_node *n1, ir_node *n2)
949 assert(get_nodes_block(n1) == get_nodes_block(n2));
951 return heights_reachable_in_block(ir_heights, n1, n2);
954 static int cmp_call_dependency(const void *c1, const void *c2)
956 ir_node *n1 = *(ir_node **) c1;
957 ir_node *n2 = *(ir_node **) c2;
961 Classical qsort() comparison function behavior:
962 0 if both elements are equal
963 1 if second is "smaller" that first
964 -1 if first is "smaller" that second
966 if (dependent_on(n1, n2))
969 if (dependent_on(n2, n1))
972 /* The nodes have no depth order, but we need a total order because qsort()
975 * Additionally, we need to respect transitive dependencies. Consider a
976 * Call a depending on Call b and an independent Call c.
977 * We MUST NOT order c > a and b > c. */
978 h1 = get_irn_height(ir_heights, n1);
979 h2 = get_irn_height(ir_heights, n2);
980 if (h1 < h2) return -1;
981 if (h1 > h2) return 1;
982 /* Same height, so use a random (but stable) order */
983 return get_irn_idx(n1) - get_irn_idx(n2);
987 * Walker: links all Call/Alloc/Free nodes to the Block they are contained.
989 static void link_ops_in_block_walker(ir_node *irn, void *data)
991 be_abi_irg_t *env = (be_abi_irg_t*)data;
992 unsigned code = get_irn_opcode(irn);
994 if (code == iro_Call ||
995 (code == iro_Alloc && get_Alloc_where(irn) == stack_alloc) ||
996 (code == iro_Free && get_Free_where(irn) == stack_alloc)) {
997 ir_node *bl = get_nodes_block(irn);
998 void *save = get_irn_link(bl);
1000 set_irn_link(irn, save);
1001 set_irn_link(bl, irn);
1004 if (code == iro_Builtin && get_Builtin_kind(irn) == ir_bk_return_address) {
1005 ir_node *param = get_Builtin_param(irn, 0);
1006 ir_tarval *tv = get_Const_tarval(param);
1007 unsigned long value = get_tarval_long(tv);
1008 /* use ebp, so the climbframe algo works... */
1010 env->call->flags.bits.try_omit_fp = 0;
1017 * Process all Call/Alloc/Free nodes inside a basic block.
1018 * Note that the link field of the block must contain a linked list of all
1019 * nodes inside the Block. We first order this list according to data dependency
1020 * and that connect the nodes together.
1022 static void process_ops_in_block(ir_node *bl, void *data)
1024 be_abi_irg_t *env = (be_abi_irg_t*)data;
1025 ir_node *curr_sp = env->init_sp;
1032 for (irn = (ir_node*)get_irn_link(bl); irn != NULL;
1033 irn = (ir_node*)get_irn_link(irn)) {
1037 nodes = ALLOCAN(ir_node*, n_nodes);
1038 for (irn = (ir_node*)get_irn_link(bl), n = 0; irn != NULL;
1039 irn = (ir_node*)get_irn_link(irn), ++n) {
1043 /* If there were call nodes in the block. */
1048 /* order the call nodes according to data dependency */
1049 qsort(nodes, n_nodes, sizeof(nodes[0]), cmp_call_dependency);
1051 for (i = n_nodes - 1; i >= 0; --i) {
1052 ir_node *irn = nodes[i];
1054 DBG((dbg, LEVEL_3, "\tprocessing call %+F\n", irn));
1055 switch (get_irn_opcode(irn)) {
1058 /* The stack pointer will be modified due to a call. */
1059 env->call->flags.bits.try_omit_fp = 0;
1061 curr_sp = adjust_call(env, irn, curr_sp);
1064 if (get_Alloc_where(irn) == stack_alloc)
1065 curr_sp = adjust_alloc(env, irn, curr_sp);
1068 if (get_Free_where(irn) == stack_alloc)
1069 curr_sp = adjust_free(env, irn, curr_sp);
1072 panic("invalid call");
1076 /* Keep the last stack state in the block by tying it to Keep node,
1077 * the proj from calls is already kept */
1078 if (curr_sp != env->init_sp &&
1079 !(is_Proj(curr_sp) && be_is_Call(get_Proj_pred(curr_sp)))) {
1081 keep = be_new_Keep(bl, 1, nodes);
1082 pmap_insert(env->keep_map, bl, keep);
1086 set_irn_link(bl, curr_sp);
1090 * Adjust all call nodes in the graph to the ABI conventions.
1092 static void process_calls(ir_graph *irg)
1094 be_abi_irg_t *abi = be_get_irg_abi(irg);
1096 irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, abi);
1098 ir_heights = heights_new(irg);
1099 irg_block_walk_graph(irg, NULL, process_ops_in_block, abi);
1100 heights_free(ir_heights);
1104 * Computes the stack argument layout type.
1105 * Changes a possibly allocated value param type by moving
1106 * entities to the stack layout type.
1108 * @param call the current call ABI
1109 * @param method_type the method type
1110 * @param param_map an array mapping method arguments to the stack layout
1113 * @return the stack argument layout type
1115 static ir_type *compute_arg_type(ir_graph *irg, be_abi_call_t *call,
1116 ir_type *method_type, ir_entity ***param_map)
1118 struct obstack *obst = be_get_be_obst(irg);
1119 ir_type *frame_type = get_irg_frame_type(irg);
1120 size_t n_params = get_method_n_params(method_type);
1121 size_t n_frame_members = get_compound_n_members(frame_type);
1122 ir_entity *va_start_entity = NULL;
1130 *param_map = map = OALLOCNZ(obst, ir_entity*, n_params);
1131 res = new_type_struct(new_id_from_chars("arg_type", 8));
1133 /* collect existing entities for value_param_types */
1134 for (f = n_frame_members; f > 0; ) {
1135 ir_entity *entity = get_compound_member(frame_type, --f);
1138 set_entity_link(entity, NULL);
1139 if (!is_parameter_entity(entity))
1141 num = get_entity_parameter_number(entity);
1142 if (num == IR_VA_START_PARAMETER_NUMBER) {
1143 /* move entity to new arg_type */
1144 set_entity_owner(entity, res);
1145 va_start_entity = entity;
1148 assert(num < n_params);
1149 if (map[num] != NULL)
1150 panic("multiple entities for parameter %u in %+F found", f, irg);
1152 if (num != n_params && !get_call_arg(call, 0, num, 1)->on_stack) {
1153 /* don't move this entity */
1158 /* move entity to new arg_type */
1159 set_entity_owner(entity, res);
1162 for (i = 0; i < n_params; ++i) {
1163 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1164 ir_type *param_type = get_method_param_type(method_type, i);
1167 if (!arg->on_stack) {
1171 if (entity == NULL) {
1172 /* create a new entity */
1173 entity = new_parameter_entity(res, i, param_type);
1175 ofs += arg->space_before;
1176 ofs = round_up2(ofs, arg->alignment);
1177 set_entity_offset(entity, ofs);
1178 ofs += arg->space_after;
1179 ofs += get_type_size_bytes(param_type);
1180 arg->stack_ent = entity;
1182 if (va_start_entity != NULL) {
1183 set_entity_offset(va_start_entity, ofs);
1185 set_type_size_bytes(res, ofs);
1186 set_type_state(res, layout_fixed);
1192 const arch_register_t *reg;
1196 static int cmp_regs(const void *a, const void *b)
1198 const reg_node_map_t *p = (const reg_node_map_t*)a;
1199 const reg_node_map_t *q = (const reg_node_map_t*)b;
1201 if (p->reg->reg_class == q->reg->reg_class)
1202 return p->reg->index - q->reg->index;
1204 return p->reg->reg_class < q->reg->reg_class ? -1 : +1;
1207 static void reg_map_to_arr(reg_node_map_t *res, pmap *reg_map)
1210 size_t n = pmap_count(reg_map);
1213 foreach_pmap(reg_map, ent) {
1214 res[i].reg = (const arch_register_t*)ent->key;
1215 res[i].irn = (ir_node*)ent->value;
1219 qsort(res, n, sizeof(res[0]), cmp_regs);
1223 * Creates a be_Return for a Return node.
1225 * @param @env the abi environment
1226 * @param irn the Return node or NULL if there was none
1227 * @param bl the block where the be_Retun should be placed
1228 * @param mem the current memory
1229 * @param n_res number of return results
1231 static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl,
1232 ir_node *mem, int n_res)
1234 be_abi_call_t *call = env->call;
1235 ir_graph *irg = get_Block_irg(bl);
1236 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1238 pmap *reg_map = pmap_create();
1239 ir_node *keep = (ir_node*)pmap_get(env->keep_map, bl);
1246 const arch_register_t **regs;
1250 get the valid stack node in this block.
1251 If we had a call in that block there is a Keep constructed by process_calls()
1252 which points to the last stack modification in that block. we'll use
1253 it then. Else we use the stack from the start block and let
1254 the ssa construction fix the usage.
1256 stack = be_abi_reg_map_get(env->regs, arch_env->sp);
1258 stack = get_irn_n(keep, 0);
1260 remove_End_keepalive(get_irg_end(irg), keep);
1263 /* Insert results for Return into the register map. */
1264 for (i = 0; i < n_res; ++i) {
1265 ir_node *res = get_Return_res(irn, i);
1266 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1267 assert(arg->in_reg && "return value must be passed in register");
1268 pmap_insert(reg_map, (void *) arg->reg, res);
1271 /* Add uses of the callee save registers. */
1272 foreach_pmap(env->regs, ent) {
1273 const arch_register_t *reg = (const arch_register_t*)ent->key;
1274 if ((reg->type & arch_register_type_ignore) || arch_register_is_callee_save(arch_env, reg))
1275 pmap_insert(reg_map, ent->key, ent->value);
1278 be_abi_reg_map_set(reg_map, arch_env->sp, stack);
1281 Maximum size of the in array for Return nodes is
1282 return args + callee save/ignore registers + memory + stack pointer
1284 in_max = pmap_count(reg_map) + n_res + 2;
1286 in = ALLOCAN(ir_node*, in_max);
1287 regs = ALLOCAN(arch_register_t const*, in_max);
1290 in[1] = be_abi_reg_map_get(reg_map, arch_env->sp);
1292 regs[1] = arch_env->sp;
1295 /* clear SP entry, since it has already been grown. */
1296 pmap_insert(reg_map, (void *) arch_env->sp, NULL);
1297 for (i = 0; i < n_res; ++i) {
1298 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1300 in[n] = be_abi_reg_map_get(reg_map, arg->reg);
1301 regs[n++] = arg->reg;
1303 /* Clear the map entry to mark the register as processed. */
1304 be_abi_reg_map_set(reg_map, arg->reg, NULL);
1307 /* grow the rest of the stuff. */
1308 foreach_pmap(reg_map, ent) {
1310 in[n] = (ir_node*)ent->value;
1311 regs[n++] = (const arch_register_t*)ent->key;
1315 /* The in array for the new back end return is now ready. */
1317 dbgi = get_irn_dbg_info(irn);
1321 /* we have to pop the shadow parameter in in case of struct returns */
1323 ret = be_new_Return(dbgi, irg, bl, n_res, pop, n, in);
1325 /* Set the register classes of the return's parameter accordingly. */
1326 for (i = 0; i < n; ++i) {
1327 if (regs[i] == NULL)
1330 be_set_constr_single_reg_in(ret, i, regs[i], arch_register_req_type_none);
1333 /* Free the space of the Epilog's in array and the register <-> proj map. */
1334 pmap_destroy(reg_map);
1339 typedef struct lower_frame_sels_env_t {
1340 ir_node *frame; /**< the current frame */
1341 const arch_register_class_t *sp_class; /**< register class of the stack pointer */
1342 const arch_register_class_t *link_class; /**< register class of the link pointer */
1343 ir_type *frame_tp; /**< the frame type */
1344 int static_link_pos; /**< argument number of the hidden static link */
1345 } lower_frame_sels_env_t;
1348 * Walker: Replaces Sels of frame type and
1349 * value param type entities by FrameAddress.
1350 * Links all used entities.
1352 static void lower_frame_sels_walker(ir_node *irn, void *data)
1354 lower_frame_sels_env_t *ctx = (lower_frame_sels_env_t*)data;
1357 ir_node *ptr = get_Sel_ptr(irn);
1359 if (ptr == ctx->frame) {
1360 ir_entity *ent = get_Sel_entity(irn);
1361 ir_node *bl = get_nodes_block(irn);
1364 nw = be_new_FrameAddr(ctx->sp_class, bl, ctx->frame, ent);
1371 * The start block has no jump, instead it has an initial exec Proj.
1372 * The backend wants to handle all blocks the same way, so we replace
1373 * the out cfg edge with a real jump.
1375 static void fix_start_block(ir_graph *irg)
1377 ir_node *initial_X = get_irg_initial_exec(irg);
1378 ir_node *start_block = get_irg_start_block(irg);
1379 ir_node *jmp = new_r_Jmp(start_block);
1381 assert(is_Proj(initial_X));
1382 exchange(initial_X, jmp);
1383 set_irg_initial_exec(irg, new_r_Bad(irg, mode_X));
1385 /* merge start block with successor if possible */
1387 const ir_edge_t *edge;
1388 foreach_out_edge(jmp, edge) {
1389 ir_node *succ = get_edge_src_irn(edge);
1390 if (!is_Block(succ))
1393 if (get_irn_arity(succ) == 1) {
1394 exchange(succ, start_block);
1402 * Modify the irg itself and the frame type.
1404 static void modify_irg(ir_graph *irg)
1406 be_abi_irg_t *env = be_get_irg_abi(irg);
1407 be_abi_call_t *call = env->call;
1408 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1409 const arch_register_t *sp = arch_env->sp;
1410 ir_type *method_type = get_entity_type(get_irg_entity(irg));
1411 be_irg_t *birg = be_birg_from_irg(irg);
1412 struct obstack *obst = be_get_be_obst(irg);
1413 be_stack_layout_t *stack_layout = be_get_irg_stack_layout(irg);
1416 ir_node *new_mem_proj;
1422 unsigned frame_size;
1425 const arch_register_t *fp_reg;
1426 ir_node *frame_pointer;
1430 const ir_edge_t *edge;
1431 ir_type *arg_type, *bet_type;
1432 lower_frame_sels_env_t ctx;
1433 ir_entity **param_map;
1435 DBG((dbg, LEVEL_1, "introducing abi on %+F\n", irg));
1437 old_mem = get_irg_initial_mem(irg);
1439 irp_reserve_resources(irp, IRP_RESOURCE_ENTITY_LINK);
1441 arg_type = compute_arg_type(irg, call, method_type, ¶m_map);
1443 /* Convert the Sel nodes in the irg to frame addr nodes: */
1444 ctx.frame = get_irg_frame(irg);
1445 ctx.sp_class = arch_env->sp->reg_class;
1446 ctx.link_class = arch_env->link_class;
1447 ctx.frame_tp = get_irg_frame_type(irg);
1449 /* layout the stackframe now */
1450 if (get_type_state(ctx.frame_tp) == layout_undefined) {
1451 default_layout_compound_type(ctx.frame_tp);
1454 /* align stackframe to 4 byte */
1455 frame_size = get_type_size_bytes(ctx.frame_tp);
1456 if (frame_size % 4 != 0) {
1457 set_type_size_bytes(ctx.frame_tp, frame_size + 4 - (frame_size % 4));
1460 env->regs = pmap_create();
1462 n_params = get_method_n_params(method_type);
1463 args = OALLOCNZ(obst, ir_node*, n_params);
1465 be_add_parameter_entity_stores(irg);
1467 irg_walk_graph(irg, lower_frame_sels_walker, NULL, &ctx);
1469 irp_free_resources(irp, IRP_RESOURCE_ENTITY_LINK);
1471 /* Fill the argument vector */
1472 arg_tuple = get_irg_args(irg);
1473 foreach_out_edge(arg_tuple, edge) {
1474 ir_node *irn = get_edge_src_irn(edge);
1475 if (! is_Anchor(irn)) {
1476 int nr = get_Proj_proj(irn);
1478 DBG((dbg, LEVEL_2, "\treading arg: %d -> %+F\n", nr, irn));
1482 stack_layout->sp_relative = call->flags.bits.try_omit_fp;
1483 bet_type = call->cb->get_between_type(irg);
1484 stack_frame_init(stack_layout, arg_type, bet_type,
1485 get_irg_frame_type(irg), param_map);
1487 /* Count the register params and add them to the number of Projs for the RegParams node */
1488 for (i = 0; i < n_params; ++i) {
1489 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1490 if (arg->in_reg && args[i]) {
1491 assert(arg->reg != sp && "cannot use stack pointer as parameter register");
1492 assert(i == get_Proj_proj(args[i]));
1494 /* For now, associate the register with the old Proj from Start representing that argument. */
1495 pmap_insert(env->regs, (void *) arg->reg, args[i]);
1496 DBG((dbg, LEVEL_2, "\targ #%d -> reg %s\n", i, arg->reg->name));
1500 /* Collect all callee-save registers */
1501 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
1502 const arch_register_class_t *cls = &arch_env->register_classes[i];
1503 for (j = 0; j < cls->n_regs; ++j) {
1504 const arch_register_t *reg = &cls->regs[j];
1505 if ((reg->type & arch_register_type_state) || arch_register_is_callee_save(arch_env, reg)) {
1506 pmap_insert(env->regs, (void *) reg, NULL);
1511 fp_reg = call->flags.bits.try_omit_fp ? arch_env->sp : arch_env->bp;
1512 rbitset_clear(birg->allocatable_regs, fp_reg->global_index);
1514 /* handle start block here (place a jump in the block) */
1515 fix_start_block(irg);
1517 pmap_insert(env->regs, (void *) sp, NULL);
1518 pmap_insert(env->regs, (void *) arch_env->bp, NULL);
1519 start_bl = get_irg_start_block(irg);
1520 env->start = be_new_Start(NULL, start_bl, pmap_count(env->regs) + 1);
1521 set_irg_start(irg, env->start);
1524 * make proj nodes for the callee save registers.
1525 * memorize them, since Return nodes get those as inputs.
1527 * Note, that if a register corresponds to an argument, the regs map
1528 * contains the old Proj from start for that argument.
1530 rm = ALLOCAN(reg_node_map_t, pmap_count(env->regs));
1531 reg_map_to_arr(rm, env->regs);
1532 for (i = 0, n = pmap_count(env->regs); i < n; ++i) {
1533 const arch_register_t *reg = rm[i].reg;
1534 ir_mode *mode = reg->reg_class->mode;
1536 arch_register_req_type_t add_type = arch_register_req_type_none;
1540 add_type |= arch_register_req_type_produces_sp;
1541 if (!rbitset_is_set(birg->allocatable_regs, reg->global_index)) {
1542 add_type |= arch_register_req_type_ignore;
1546 proj = new_r_Proj(env->start, mode, nr + 1);
1547 pmap_insert(env->regs, (void *) reg, proj);
1548 be_set_constr_single_reg_out(env->start, nr + 1, reg, add_type);
1549 arch_set_irn_register(proj, reg);
1551 DBG((dbg, LEVEL_2, "\tregister save proj #%d -> reg %s\n", nr, reg->name));
1554 /* create a new initial memory proj */
1555 assert(is_Proj(old_mem));
1556 arch_set_irn_register_req_out(env->start, 0, arch_no_register_req);
1557 new_mem_proj = new_r_Proj(env->start, mode_M, 0);
1559 set_irg_initial_mem(irg, mem);
1561 env->init_sp = be_abi_reg_map_get(env->regs, sp);
1563 /* set new frame_pointer */
1564 frame_pointer = be_abi_reg_map_get(env->regs, fp_reg);
1565 set_irg_frame(irg, frame_pointer);
1567 /* rewire old mem users to new mem */
1568 exchange(old_mem, mem);
1570 /* keep the mem (for functions with an endless loop = no return) */
1573 set_irg_initial_mem(irg, mem);
1575 /* Now, introduce stack param nodes for all parameters passed on the stack */
1576 for (i = 0; i < n_params; ++i) {
1577 ir_node *arg_proj = args[i];
1578 ir_node *repl = NULL;
1580 if (arg_proj != NULL) {
1581 be_abi_call_arg_t *arg;
1582 ir_type *param_type;
1583 int nr = get_Proj_proj(arg_proj);
1586 nr = MIN(nr, n_params);
1587 arg = get_call_arg(call, 0, nr, 1);
1588 param_type = get_method_param_type(method_type, nr);
1591 repl = (ir_node*)pmap_get(env->regs, arg->reg);
1592 } else if (arg->on_stack) {
1593 ir_node *addr = be_new_FrameAddr(sp->reg_class, start_bl, frame_pointer, arg->stack_ent);
1595 /* For atomic parameters which are actually used, we create a Load node. */
1596 if (is_atomic_type(param_type) && get_irn_n_edges(args[i]) > 0) {
1597 ir_mode *mode = get_type_mode(param_type);
1598 ir_mode *load_mode = arg->load_mode;
1599 ir_node *nomem = get_irg_no_mem(irg);
1601 ir_node *load = new_r_Load(start_bl, nomem, addr, load_mode, cons_floats);
1602 repl = new_r_Proj(load, load_mode, pn_Load_res);
1604 if (mode != load_mode) {
1605 repl = new_r_Conv(start_bl, repl, mode);
1608 /* The stack parameter is not primitive (it is a struct or array),
1609 * we thus will create a node representing the parameter's address
1615 assert(repl != NULL);
1617 /* Beware: the mode of the register parameters is always the mode of the register class
1618 which may be wrong. Add Conv's then. */
1619 mode = get_irn_mode(args[i]);
1620 if (mode != get_irn_mode(repl)) {
1621 repl = new_r_Conv(get_nodes_block(repl), repl, mode);
1623 exchange(args[i], repl);
1627 /* the arg proj is not needed anymore now and should be only used by the anchor */
1628 assert(get_irn_n_edges(arg_tuple) == 1);
1629 kill_node(arg_tuple);
1630 set_irg_args(irg, new_r_Bad(irg, mode_T));
1632 /* All Return nodes hang on the End node, so look for them there. */
1633 end = get_irg_end_block(irg);
1634 for (i = 0, n = get_Block_n_cfgpreds(end); i < n; ++i) {
1635 ir_node *irn = get_Block_cfgpred(end, i);
1637 if (is_Return(irn)) {
1638 ir_node *blk = get_nodes_block(irn);
1639 ir_node *mem = get_Return_mem(irn);
1640 ir_node *ret = create_be_return(env, irn, blk, mem, get_Return_n_ress(irn));
1645 /* if we have endless loops here, n might be <= 0. Do NOT create a be_Return then,
1646 the code is dead and will never be executed. */
1649 /** Fix the state inputs of calls that still hang on unknowns */
1650 static void fix_call_state_inputs(ir_graph *irg)
1652 be_abi_irg_t *env = be_get_irg_abi(irg);
1653 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1655 arch_register_t **stateregs = NEW_ARR_F(arch_register_t*, 0);
1657 /* Collect caller save registers */
1658 n = arch_env->n_register_classes;
1659 for (i = 0; i < n; ++i) {
1661 const arch_register_class_t *cls = &arch_env->register_classes[i];
1662 for (j = 0; j < cls->n_regs; ++j) {
1663 const arch_register_t *reg = arch_register_for_index(cls, j);
1664 if (reg->type & arch_register_type_state) {
1665 ARR_APP1(arch_register_t*, stateregs, (arch_register_t *)reg);
1670 n = ARR_LEN(env->calls);
1671 n_states = ARR_LEN(stateregs);
1672 for (i = 0; i < n; ++i) {
1674 ir_node *call = env->calls[i];
1676 arity = get_irn_arity(call);
1678 /* the state reg inputs are the last n inputs of the calls */
1679 for (s = 0; s < n_states; ++s) {
1680 int inp = arity - n_states + s;
1681 const arch_register_t *reg = stateregs[s];
1682 ir_node *regnode = be_abi_reg_map_get(env->regs, reg);
1684 set_irn_n(call, inp, regnode);
1688 DEL_ARR_F(stateregs);
1692 * Create a trampoline entity for the given method.
1694 static ir_entity *create_trampoline(be_main_env_t *be, ir_entity *method)
1696 ir_type *type = get_entity_type(method);
1697 ident *old_id = get_entity_ld_ident(method);
1698 ident *id = id_mangle3("", old_id, "$stub");
1699 ir_type *parent = be->pic_trampolines_type;
1700 ir_entity *ent = new_entity(parent, old_id, type);
1701 set_entity_ld_ident(ent, id);
1702 set_entity_visibility(ent, ir_visibility_private);
1708 * Returns the trampoline entity for the given method.
1710 static ir_entity *get_trampoline(be_main_env_t *env, ir_entity *method)
1712 ir_entity *result = (ir_entity*)pmap_get(env->ent_trampoline_map, method);
1713 if (result == NULL) {
1714 result = create_trampoline(env, method);
1715 pmap_insert(env->ent_trampoline_map, method, result);
1721 static ir_entity *create_pic_symbol(be_main_env_t *be, ir_entity *entity)
1723 ident *old_id = get_entity_ld_ident(entity);
1724 ident *id = id_mangle3("", old_id, "$non_lazy_ptr");
1725 ir_type *e_type = get_entity_type(entity);
1726 ir_type *type = new_type_pointer(e_type);
1727 ir_type *parent = be->pic_symbols_type;
1728 ir_entity *ent = new_entity(parent, old_id, type);
1729 set_entity_ld_ident(ent, id);
1730 set_entity_visibility(ent, ir_visibility_private);
1735 static ir_entity *get_pic_symbol(be_main_env_t *env, ir_entity *entity)
1737 ir_entity *result = (ir_entity*)pmap_get(env->ent_pic_symbol_map, entity);
1738 if (result == NULL) {
1739 result = create_pic_symbol(env, entity);
1740 pmap_insert(env->ent_pic_symbol_map, entity, result);
1749 * Returns non-zero if a given entity can be accessed using a relative address.
1751 static int can_address_relative(ir_entity *entity)
1753 return get_entity_visibility(entity) != ir_visibility_external
1754 && !(get_entity_linkage(entity) & IR_LINKAGE_MERGE);
1757 static ir_node *get_pic_base(ir_graph *irg)
1759 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1760 if (arch_env->impl->get_pic_base == NULL)
1762 return arch_env->impl->get_pic_base(irg);
1765 /** patches SymConsts to work in position independent code */
1766 static void fix_pic_symconsts(ir_node *node, void *data)
1768 ir_graph *irg = get_irn_irg(node);
1769 be_main_env_t *be = be_get_irg_main_env(irg);
1779 arity = get_irn_arity(node);
1780 for (i = 0; i < arity; ++i) {
1782 ir_node *pred = get_irn_n(node, i);
1784 ir_entity *pic_symbol;
1785 ir_node *pic_symconst;
1787 if (!is_SymConst(pred))
1790 entity = get_SymConst_entity(pred);
1791 block = get_nodes_block(pred);
1793 /* calls can jump to relative addresses, so we can directly jump to
1794 the (relatively) known call address or the trampoline */
1795 if (i == 1 && is_Call(node)) {
1796 ir_entity *trampoline;
1797 ir_node *trampoline_const;
1799 if (can_address_relative(entity))
1802 dbgi = get_irn_dbg_info(pred);
1803 trampoline = get_trampoline(be, entity);
1804 trampoline_const = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
1806 set_irn_n(node, i, trampoline_const);
1810 /* everything else is accessed relative to EIP */
1811 mode = get_irn_mode(pred);
1812 pic_base = get_pic_base(irg);
1814 /* all ok now for locally constructed stuff */
1815 if (can_address_relative(entity)) {
1816 ir_node *add = new_r_Add(block, pic_base, pred, mode);
1818 /* make sure the walker doesn't visit this add again */
1819 mark_irn_visited(add);
1820 set_irn_n(node, i, add);
1824 /* get entry from pic symbol segment */
1825 dbgi = get_irn_dbg_info(pred);
1826 pic_symbol = get_pic_symbol(be, entity);
1827 pic_symconst = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
1829 add = new_r_Add(block, pic_base, pic_symconst, mode);
1830 mark_irn_visited(add);
1832 /* we need an extra indirection for global data outside our current
1833 module. The loads are always safe and can therefore float
1834 and need no memory input */
1835 load = new_r_Load(block, get_irg_no_mem(irg), add, mode, cons_floats);
1836 load_res = new_r_Proj(load, mode, pn_Load_res);
1838 set_irn_n(node, i, load_res);
1842 void be_abi_introduce(ir_graph *irg)
1844 be_abi_irg_t *env = XMALLOCZ(be_abi_irg_t);
1845 ir_node *old_frame = get_irg_frame(irg);
1846 be_options_t *options = be_get_irg_options(irg);
1847 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1848 ir_entity *entity = get_irg_entity(irg);
1849 ir_type *method_type = get_entity_type(entity);
1850 be_irg_t *birg = be_birg_from_irg(irg);
1851 struct obstack *obst = &birg->obst;
1852 ir_node *dummy = new_r_Dummy(irg,
1853 arch_env->sp->reg_class->mode);
1856 /* determine allocatable registers */
1857 assert(birg->allocatable_regs == NULL);
1858 birg->allocatable_regs = rbitset_obstack_alloc(obst, arch_env->n_registers);
1859 for (r = 0; r < arch_env->n_registers; ++r) {
1860 const arch_register_t *reg = &arch_env->registers[r];
1861 if ( !(reg->type & arch_register_type_ignore)) {
1862 rbitset_set(birg->allocatable_regs, r);
1866 /* break here if backend provides a custom API.
1867 * Note: we shouldn't have to setup any be_abi_irg_t* stuff at all,
1868 * but need more cleanup to make this work
1870 be_set_irg_abi(irg, env);
1872 be_omit_fp = options->omit_fp;
1874 env->keep_map = pmap_create();
1875 env->call = be_abi_call_new(arch_env->sp->reg_class);
1876 arch_env_get_call_abi(arch_env, method_type, env->call);
1878 env->init_sp = dummy;
1879 env->calls = NEW_ARR_F(ir_node*, 0);
1884 irg_walk_graph(irg, fix_pic_symconsts, NULL, env);
1887 /* Lower all call nodes in the IRG. */
1890 /* Process the IRG */
1893 /* fix call inputs for state registers */
1894 fix_call_state_inputs(irg);
1896 /* We don't need the keep map anymore. */
1897 pmap_destroy(env->keep_map);
1898 env->keep_map = NULL;
1900 /* calls array is not needed anymore */
1901 DEL_ARR_F(env->calls);
1904 /* reroute the stack origin of the calls to the true stack origin. */
1905 exchange(dummy, env->init_sp);
1906 exchange(old_frame, get_irg_frame(irg));
1908 pmap_destroy(env->regs);
1912 void be_abi_free(ir_graph *irg)
1914 be_abi_irg_t *env = be_get_irg_abi(irg);
1916 if (env->call != NULL)
1917 be_abi_call_free(env->call);
1918 assert(env->regs == NULL);
1921 be_set_irg_abi(irg, NULL);
1924 void be_put_allocatable_regs(const ir_graph *irg,
1925 const arch_register_class_t *cls, bitset_t *bs)
1927 be_irg_t *birg = be_birg_from_irg(irg);
1928 unsigned *allocatable_regs = birg->allocatable_regs;
1931 assert(bitset_size(bs) == cls->n_regs);
1932 bitset_clear_all(bs);
1933 for (i = 0; i < cls->n_regs; ++i) {
1934 const arch_register_t *reg = &cls->regs[i];
1935 if (rbitset_is_set(allocatable_regs, reg->global_index))
1940 unsigned be_get_n_allocatable_regs(const ir_graph *irg,
1941 const arch_register_class_t *cls)
1943 bitset_t *bs = bitset_alloca(cls->n_regs);
1944 be_put_allocatable_regs(irg, cls, bs);
1945 return bitset_popcount(bs);
1948 void be_set_allocatable_regs(const ir_graph *irg,
1949 const arch_register_class_t *cls,
1950 unsigned *raw_bitset)
1952 be_irg_t *birg = be_birg_from_irg(irg);
1953 unsigned *allocatable_regs = birg->allocatable_regs;
1956 rbitset_clear_all(raw_bitset, cls->n_regs);
1957 for (i = 0; i < cls->n_regs; ++i) {
1958 const arch_register_t *reg = &cls->regs[i];
1959 if (rbitset_is_set(allocatable_regs, reg->global_index))
1960 rbitset_set(raw_bitset, i);
1964 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_abi)
1965 void be_init_abi(void)
1967 FIRM_DBG_REGISTER(dbg, "firm.be.abi");