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 pmap *regs; /**< A map of all callee-save and ignore regs to
95 their Projs to the RegParams node. */
96 pmap *keep_map; /**< mapping blocks to keep nodes. */
98 ir_node **calls; /**< flexible array containing all be_Call nodes */
101 static ir_heights_t *ir_heights;
103 /** Flag: if set, try to omit the frame pointer in all routines. */
104 static int be_omit_fp = 1;
106 static ir_node *be_abi_reg_map_get(pmap *map, const arch_register_t *reg)
108 return pmap_get(ir_node, map, reg);
111 static void be_abi_reg_map_set(pmap *map, const arch_register_t* reg,
114 pmap_insert(map, reg, node);
118 * Check if the given register is callee save, ie. will be saved by the callee.
120 static bool arch_register_is_callee_save(
121 const arch_env_t *arch_env,
122 const arch_register_t *reg)
124 if (arch_env->impl->register_saved_by)
125 return arch_env->impl->register_saved_by(reg, /*callee=*/1);
130 * Check if the given register is caller save, ie. must be saved by the caller.
132 static bool arch_register_is_caller_save(
133 const arch_env_t *arch_env,
134 const arch_register_t *reg)
136 if (arch_env->impl->register_saved_by)
137 return arch_env->impl->register_saved_by(reg, /*callee=*/0);
144 _ ____ ___ ____ _ _ _ _
145 / \ | __ )_ _| / ___|__ _| | | |__ __ _ ___| | _____
146 / _ \ | _ \| | | | / _` | | | '_ \ / _` |/ __| |/ / __|
147 / ___ \| |_) | | | |__| (_| | | | |_) | (_| | (__| <\__ \
148 /_/ \_\____/___| \____\__,_|_|_|_.__/ \__,_|\___|_|\_\___/
150 These callbacks are used by the backend to set the parameters
151 for a specific call type.
155 * Set compare function: compares two ABI call object arguments.
157 static int cmp_call_arg(const void *a, const void *b, size_t n)
159 const be_abi_call_arg_t *p = (const be_abi_call_arg_t*)a;
160 const be_abi_call_arg_t *q = (const be_abi_call_arg_t*)b;
162 return !(p->is_res == q->is_res && p->pos == q->pos && p->callee == q->callee);
166 * Get an ABI call object argument.
168 * @param call the abi call
169 * @param is_res true for call results, false for call arguments
170 * @param pos position of the argument
171 * @param callee context type - if we are callee or caller
173 static be_abi_call_arg_t *get_call_arg(be_abi_call_t *call, int is_res, int pos, int callee)
175 be_abi_call_arg_t arg;
178 memset(&arg, 0, sizeof(arg));
183 hash = is_res * 128 + pos;
185 return set_find(be_abi_call_arg_t, call->params, &arg, sizeof(arg), hash);
189 * Set an ABI call object argument.
191 static void remember_call_arg(be_abi_call_arg_t *arg, be_abi_call_t *call, be_abi_context_t context)
193 unsigned hash = arg->is_res * 128 + arg->pos;
194 if (context & ABI_CONTEXT_CALLEE) {
196 (void)set_insert(be_abi_call_arg_t, call->params, arg, sizeof(*arg), hash);
198 if (context & ABI_CONTEXT_CALLER) {
200 (void)set_insert(be_abi_call_arg_t, call->params, arg, sizeof(*arg), hash);
204 /* Set the flags for a call. */
205 void be_abi_call_set_flags(be_abi_call_t *call, be_abi_call_flags_t flags, const be_abi_callbacks_t *cb)
211 /* Sets the number of bytes the stackframe is shrinked by the callee on return */
212 void be_abi_call_set_pop(be_abi_call_t *call, int pop)
218 /* Set register class for call address */
219 void be_abi_call_set_call_address_reg_class(be_abi_call_t *call, const arch_register_class_t *cls)
221 call->cls_addr = cls;
225 void be_abi_call_param_stack(be_abi_call_t *call, int arg_pos,
226 ir_mode *load_mode, unsigned alignment,
227 unsigned space_before, unsigned space_after,
228 be_abi_context_t context)
230 be_abi_call_arg_t arg;
231 memset(&arg, 0, sizeof(arg));
232 assert(alignment > 0 && "Alignment must be greater than 0");
234 arg.load_mode = load_mode;
235 arg.alignment = alignment;
236 arg.space_before = space_before;
237 arg.space_after = space_after;
241 remember_call_arg(&arg, call, context);
244 void be_abi_call_param_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
246 be_abi_call_arg_t arg;
247 memset(&arg, 0, sizeof(arg));
254 remember_call_arg(&arg, call, context);
257 void be_abi_call_res_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
259 be_abi_call_arg_t arg;
260 memset(&arg, 0, sizeof(arg));
267 remember_call_arg(&arg, call, context);
270 /* Get the flags of a ABI call object. */
271 be_abi_call_flags_t be_abi_call_get_flags(const be_abi_call_t *call)
277 * Constructor for a new ABI call object.
279 * @param cls_addr register class of the call address
281 * @return the new ABI call object
283 static be_abi_call_t *be_abi_call_new(const arch_register_class_t *cls_addr)
285 be_abi_call_t *call = XMALLOCZ(be_abi_call_t);
288 call->params = new_set(cmp_call_arg, 16);
290 call->cls_addr = cls_addr;
292 call->flags.bits.try_omit_fp = be_omit_fp;
298 * Destructor for an ABI call object.
300 static void be_abi_call_free(be_abi_call_t *call)
302 del_set(call->params);
307 * Initializes the frame layout from parts
309 * @param frame the stack layout that will be initialized
310 * @param args the stack argument layout type
311 * @param between the between layout type
312 * @param locals the method frame type
314 * @return the initialized stack layout
316 static be_stack_layout_t *stack_frame_init(be_stack_layout_t *frame, ir_type *args,
317 ir_type *between, ir_type *locals)
319 frame->arg_type = args;
320 frame->between_type = between;
321 frame->frame_type = locals;
322 frame->initial_offset = 0;
323 frame->initial_bias = 0;
324 frame->order[1] = between;
326 /* typical decreasing stack: locals have the
327 * lowest addresses, arguments the highest */
328 frame->order[0] = locals;
329 frame->order[2] = args;
340 Adjustment of the calls inside a graph.
345 * Transform a call node into a be_Call node.
347 * @param env The ABI environment for the current irg.
348 * @param irn The call node.
349 * @param curr_sp The stack pointer node to use.
350 * @return The stack pointer after the call.
352 static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
354 ir_graph *irg = get_irn_irg(irn);
355 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
356 ir_type *call_tp = get_Call_type(irn);
357 ir_node *call_ptr = get_Call_ptr(irn);
358 size_t n_params = get_method_n_params(call_tp);
359 ir_node *curr_mem = get_Call_mem(irn);
360 ir_node *bl = get_nodes_block(irn);
362 const arch_register_t *sp = arch_env->sp;
363 be_abi_call_t *call = be_abi_call_new(sp->reg_class);
364 ir_mode *mach_mode = sp->reg_class->mode;
365 int n_res = get_method_n_ress(call_tp);
367 ir_node *res_proj = NULL;
368 int n_reg_params = 0;
369 int n_stack_params = 0;
372 const arch_register_t **states = NEW_ARR_F(const arch_register_t*, 0);
373 const arch_register_t **destroyed_regs = NEW_ARR_F(const arch_register_t*, 0);
377 int n_reg_results = 0;
379 int *stack_param_idx;
381 int throws_exception;
386 /* Let the isa fill out the abi description for that call node. */
387 arch_env_get_call_abi(arch_env, call_tp, call);
389 /* Insert code to put the stack arguments on the stack. */
390 assert((size_t)get_Call_n_params(irn) == n_params);
391 stack_param_idx = ALLOCAN(int, n_params);
392 for (p = 0; p < n_params; ++p) {
393 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
396 int arg_size = get_type_size_bytes(get_method_param_type(call_tp, p));
398 stack_size += round_up2(arg->space_before, arg->alignment);
399 stack_size += round_up2(arg_size, arg->alignment);
400 stack_size += round_up2(arg->space_after, arg->alignment);
402 stack_param_idx[n_stack_params++] = p;
406 /* Collect all arguments which are passed in registers. */
407 reg_param_idxs = ALLOCAN(int, n_params);
408 for (p = 0; p < n_params; ++p) {
409 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
410 if (arg && arg->in_reg) {
411 reg_param_idxs[n_reg_params++] = p;
416 * If the stack is decreasing and we do not want to store sequentially,
417 * or someone else allocated the call frame
418 * we allocate as much space on the stack all parameters need, by
419 * moving the stack pointer along the stack's direction.
421 * Note: we also have to do this for stack_size == 0, because we may have
422 * to adjust stack alignment for the call.
424 curr_sp = be_new_IncSP(sp, bl, curr_sp, stack_size, 1);
426 dbgi = get_irn_dbg_info(irn);
427 /* If there are some parameters which shall be passed on the stack. */
428 if (n_stack_params > 0) {
430 ir_node **in = ALLOCAN(ir_node*, n_stack_params+1);
433 curr_mem = get_Call_mem(irn);
434 in[n_in++] = curr_mem;
436 for (i = 0; i < n_stack_params; ++i) {
437 int p = stack_param_idx[i];
438 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
439 ir_node *param = get_Call_param(irn, p);
440 ir_node *addr = curr_sp;
442 ir_type *param_type = get_method_param_type(call_tp, p);
443 int param_size = get_type_size_bytes(param_type) + arg->space_after;
446 * If we wanted to build the arguments sequentially,
447 * the stack pointer for the next must be incremented,
448 * and the memory value propagated.
450 curr_ofs += arg->space_before;
451 curr_ofs = round_up2(curr_ofs, arg->alignment);
453 /* Make the expression to compute the argument's offset. */
455 ir_mode *constmode = mach_mode;
456 if (mode_is_reference(mach_mode)) {
459 addr = new_r_Const_long(irg, constmode, curr_ofs);
460 addr = new_r_Add(bl, curr_sp, addr, mach_mode);
463 /* Insert a store for primitive arguments. */
464 if (is_atomic_type(param_type)) {
465 ir_node *nomem = get_irg_no_mem(irg);
466 ir_node *mem_input = nomem;
467 ir_node *store = new_rd_Store(dbgi, bl, mem_input, addr, param, cons_none);
468 mem = new_r_Proj(store, mode_M, pn_Store_M);
470 /* Make a mem copy for compound arguments. */
473 assert(mode_is_reference(get_irn_mode(param)));
474 copy = new_rd_CopyB(dbgi, bl, curr_mem, addr, param, param_type);
475 mem = new_r_Proj(copy, mode_M, pn_CopyB_M);
478 curr_ofs += param_size;
483 /* We need the sync only, if we didn't build the stores sequentially. */
484 if (n_stack_params >= 1) {
485 curr_mem = new_r_Sync(bl, n_in, in);
487 curr_mem = get_Call_mem(irn);
491 /* Put caller save into the destroyed set and state registers in the states
493 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
495 const arch_register_class_t *cls = &arch_env->register_classes[i];
496 for (j = 0; j < cls->n_regs; ++j) {
497 const arch_register_t *reg = arch_register_for_index(cls, j);
499 /* even if destroyed all is specified, neither SP nor FP are
500 * destroyed (else bad things will happen) */
501 if (reg == arch_env->sp || reg == arch_env->bp)
504 if (reg->type & arch_register_type_state) {
505 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
506 ARR_APP1(const arch_register_t*, states, reg);
507 /* we're already in the destroyed set so no need for further
511 if (arch_register_is_caller_save(arch_env, reg)) {
512 if (!(reg->type & arch_register_type_ignore)) {
513 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
519 /* search the largest result proj number */
520 res_projs = ALLOCANZ(ir_node*, n_res);
522 foreach_out_edge(irn, edge) {
523 ir_node *irn = get_edge_src_irn(edge);
525 if (!is_Proj(irn) || get_Proj_proj(irn) != pn_Call_T_result)
528 foreach_out_edge(irn, res_edge) {
530 ir_node *res = get_edge_src_irn(res_edge);
532 assert(is_Proj(res));
534 proj = get_Proj_proj(res);
535 assert(proj < n_res);
536 assert(res_projs[proj] == NULL);
537 res_projs[proj] = res;
543 /** TODO: this is not correct for cases where return values are passed
544 * on the stack, but no known ABI does this currently...
546 n_reg_results = n_res;
549 in = ALLOCAN(ir_node*, n_reg_params + ARR_LEN(states));
551 /* make the back end call node and set its register requirements. */
552 for (i = 0; i < n_reg_params; ++i) {
553 in[n_ins++] = get_Call_param(irn, reg_param_idxs[i]);
556 /* add state registers ins */
557 for (s = 0; s < ARR_LEN(states); ++s) {
558 const arch_register_t *reg = states[s];
559 const arch_register_class_t *cls = reg->reg_class;
560 ir_node *regnode = new_r_Unknown(irg, cls->mode);
561 in[n_ins++] = regnode;
563 assert(n_ins == (int) (n_reg_params + ARR_LEN(states)));
565 /* ins collected, build the call */
566 throws_exception = ir_throws_exception(irn);
567 if (env->call->flags.bits.call_has_imm && is_SymConst(call_ptr)) {
569 low_call = be_new_Call(dbgi, irg, bl, curr_mem, sp->single_req, curr_sp,
570 sp->single_req, curr_sp,
571 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
572 n_ins, in, get_Call_type(irn));
573 be_Call_set_entity(low_call, get_SymConst_entity(call_ptr));
576 low_call = be_new_Call(dbgi, irg, bl, curr_mem, sp->single_req, curr_sp,
577 call->cls_addr->class_req, call_ptr,
578 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
579 n_ins, in, get_Call_type(irn));
581 ir_set_throws_exception(low_call, throws_exception);
582 be_Call_set_pop(low_call, call->pop);
584 /* put the call into the list of all calls for later processing */
585 ARR_APP1(ir_node *, env->calls, low_call);
587 /* create new stack pointer */
588 curr_sp = new_r_Proj(low_call, get_irn_mode(curr_sp), pn_be_Call_sp);
589 be_set_constr_single_reg_out(low_call, pn_be_Call_sp, sp,
590 arch_register_req_type_ignore | arch_register_req_type_produces_sp);
591 arch_set_irn_register(curr_sp, sp);
593 /* now handle results */
594 for (i = 0; i < n_res; ++i) {
595 ir_node *proj = res_projs[i];
596 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
598 /* returns values on stack not supported yet */
602 shift the proj number to the right, since we will drop the
603 unspeakable Proj_T from the Call. Therefore, all real argument
604 Proj numbers must be increased by pn_be_Call_first_res
606 long pn = i + pn_be_Call_first_res;
609 ir_type *res_type = get_method_res_type(call_tp, i);
610 ir_mode *mode = get_type_mode(res_type);
611 proj = new_r_Proj(low_call, mode, pn);
614 set_Proj_pred(proj, low_call);
615 set_Proj_proj(proj, pn);
619 /* remove register from destroyed regs */
621 size_t n = ARR_LEN(destroyed_regs);
622 for (j = 0; j < n; ++j) {
623 if (destroyed_regs[j] == arg->reg) {
624 destroyed_regs[j] = destroyed_regs[n-1];
625 ARR_SHRINKLEN(destroyed_regs,n-1);
632 DBG((dbg, LEVEL_3, "\tcreated backend call %+F\n", low_call));
634 /* Set the register classes and constraints of the Call parameters. */
635 for (i = 0; i < n_reg_params; ++i) {
636 int index = reg_param_idxs[i];
637 be_abi_call_arg_t *arg = get_call_arg(call, 0, index, 0);
638 assert(arg->reg != NULL);
640 be_set_constr_single_reg_in(low_call, n_be_Call_first_arg + i,
641 arg->reg, arch_register_req_type_none);
644 /* Set the register constraints of the results. */
645 for (i = 0; i < n_res; ++i) {
646 ir_node *proj = res_projs[i];
647 const be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
648 int pn = get_Proj_proj(proj);
651 be_set_constr_single_reg_out(low_call, pn, arg->reg,
652 arch_register_req_type_none);
653 arch_set_irn_register(proj, arg->reg);
655 exchange(irn, low_call);
657 /* kill the ProjT node */
658 if (res_proj != NULL) {
662 /* Make additional projs for the caller save registers
663 and the Keep node which keeps them alive. */
669 int curr_res_proj = pn_be_Call_first_res + n_reg_results;
672 n_ins = ARR_LEN(destroyed_regs) + n_reg_results + 1;
673 in = ALLOCAN(ir_node *, n_ins);
675 /* also keep the stack pointer */
676 set_irn_link(curr_sp, (void*) sp);
679 for (d = 0; d < ARR_LEN(destroyed_regs); ++d) {
680 const arch_register_t *reg = destroyed_regs[d];
681 ir_node *proj = new_r_Proj(low_call, reg->reg_class->mode, curr_res_proj);
683 /* memorize the register in the link field. we need afterwards to set the register class of the keep correctly. */
684 be_set_constr_single_reg_out(low_call, curr_res_proj, reg,
685 arch_register_req_type_none);
686 arch_set_irn_register(proj, reg);
688 set_irn_link(proj, (void*) reg);
693 for (i = 0; i < n_reg_results; ++i) {
694 ir_node *proj = res_projs[i];
695 const arch_register_t *reg = arch_get_irn_register(proj);
696 set_irn_link(proj, (void*) reg);
701 /* create the Keep for the caller save registers */
702 keep = be_new_Keep(bl, n, in);
703 for (i = 0; i < n; ++i) {
704 const arch_register_t *reg = (const arch_register_t*)get_irn_link(in[i]);
705 be_node_set_reg_class_in(keep, i, reg->reg_class);
709 /* Clean up the stack. */
710 assert(stack_size >= call->pop);
711 stack_size -= call->pop;
713 if (stack_size > 0) {
714 ir_node *mem_proj = NULL;
716 foreach_out_edge(low_call, edge) {
717 ir_node *irn = get_edge_src_irn(edge);
718 if (is_Proj(irn) && get_Proj_proj(irn) == pn_Call_M) {
725 mem_proj = new_r_Proj(low_call, mode_M, pn_be_Call_M);
726 keep_alive(mem_proj);
729 /* Clean up the stack frame or revert alignment fixes if we allocated it */
730 curr_sp = be_new_IncSP(sp, bl, curr_sp, -stack_size, 0);
732 be_abi_call_free(call);
735 DEL_ARR_F(destroyed_regs);
741 * Adjust the size of a node representing a stack alloc or free for the minimum stack alignment.
743 * @param alignment the minimum stack alignment
744 * @param size the node containing the non-aligned size
745 * @param block the block where new nodes are allocated on
746 * @param dbg debug info for new nodes
748 * @return a node representing the aligned size
750 static ir_node *adjust_alloc_size(unsigned stack_alignment, ir_node *size,
751 ir_node *block, dbg_info *dbg)
753 if (stack_alignment > 1) {
759 assert(is_po2(stack_alignment));
761 mode = get_irn_mode(size);
762 tv = new_tarval_from_long(stack_alignment-1, mode);
763 irg = get_Block_irg(block);
764 mask = new_r_Const(irg, tv);
765 size = new_rd_Add(dbg, block, size, mask, mode);
767 tv = new_tarval_from_long(-(long)stack_alignment, mode);
768 mask = new_r_Const(irg, tv);
769 size = new_rd_And(dbg, block, size, mask, mode);
775 * The alloca is transformed into a back end alloca node and connected to the stack nodes.
777 static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp)
779 ir_node *block = get_nodes_block(alloc);
780 ir_graph *irg = get_Block_irg(block);
781 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
782 ir_node *alloc_mem = NULL;
783 ir_node *alloc_res = NULL;
784 ir_type *type = get_Alloc_type(alloc);
791 unsigned stack_alignment;
793 /* all non-stack Alloc nodes should already be lowered before the backend */
794 assert(get_Alloc_where(alloc) == stack_alloc);
796 foreach_out_edge(alloc, edge) {
797 ir_node *irn = get_edge_src_irn(edge);
799 assert(is_Proj(irn));
800 switch (get_Proj_proj(irn)) {
812 /* Beware: currently Alloc nodes without a result might happen,
813 only escape analysis kills them and this phase runs only for object
814 oriented source. We kill the Alloc here. */
815 if (alloc_res == NULL && alloc_mem) {
816 exchange(alloc_mem, get_Alloc_mem(alloc));
820 dbg = get_irn_dbg_info(alloc);
821 count = get_Alloc_count(alloc);
823 /* we might need to multiply the count with the element size */
824 if (!is_unknown_type(type) && get_type_size_bytes(type) != 1) {
825 ir_mode *mode = get_irn_mode(count);
826 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type),
828 ir_node *cnst = new_rd_Const(dbg, irg, tv);
829 size = new_rd_Mul(dbg, block, count, cnst, mode);
834 /* The stack pointer will be modified in an unknown manner.
835 We cannot omit it. */
836 env->call->flags.bits.try_omit_fp = 0;
838 stack_alignment = 1 << arch_env->stack_alignment;
839 size = adjust_alloc_size(stack_alignment, size, block, dbg);
840 new_alloc = be_new_AddSP(arch_env->sp, block, curr_sp, size);
841 set_irn_dbg_info(new_alloc, dbg);
843 if (alloc_mem != NULL) {
847 addsp_mem = new_r_Proj(new_alloc, mode_M, pn_be_AddSP_M);
849 /* We need to sync the output mem of the AddSP with the input mem
850 edge into the alloc node. */
851 ins[0] = get_Alloc_mem(alloc);
853 sync = new_r_Sync(block, 2, ins);
855 exchange(alloc_mem, sync);
858 exchange(alloc, new_alloc);
860 /* fix projnum of alloca res */
861 set_Proj_proj(alloc_res, pn_be_AddSP_res);
863 curr_sp = new_r_Proj(new_alloc, get_irn_mode(curr_sp), pn_be_AddSP_sp);
870 * The Free is transformed into a back end free node and connected to the stack nodes.
872 static ir_node *adjust_free(be_abi_irg_t *env, ir_node *free, ir_node *curr_sp)
874 ir_node *block = get_nodes_block(free);
875 ir_graph *irg = get_irn_irg(free);
876 ir_type *type = get_Free_type(free);
877 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
878 ir_mode *sp_mode = arch_env->sp->reg_class->mode;
879 dbg_info *dbg = get_irn_dbg_info(free);
880 ir_node *subsp, *mem, *res, *size, *sync;
882 unsigned stack_alignment;
884 /* all non-stack-alloc Free nodes should already be lowered before the
886 assert(get_Free_where(free) == stack_alloc);
888 /* we might need to multiply the size with the element size */
889 if (!is_unknown_type(type) && get_type_size_bytes(type) != 1) {
890 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type), mode_Iu);
891 ir_node *cnst = new_rd_Const(dbg, irg, tv);
892 ir_node *mul = new_rd_Mul(dbg, block, get_Free_count(free),
896 size = get_Free_count(free);
899 stack_alignment = 1 << arch_env->stack_alignment;
900 size = adjust_alloc_size(stack_alignment, size, block, dbg);
902 /* The stack pointer will be modified in an unknown manner.
903 We cannot omit it. */
904 env->call->flags.bits.try_omit_fp = 0;
905 subsp = be_new_SubSP(arch_env->sp, block, curr_sp, size);
906 set_irn_dbg_info(subsp, dbg);
908 mem = new_r_Proj(subsp, mode_M, pn_be_SubSP_M);
909 res = new_r_Proj(subsp, sp_mode, pn_be_SubSP_sp);
911 /* we need to sync the memory */
912 in[0] = get_Free_mem(free);
914 sync = new_r_Sync(block, 2, in);
916 /* and make the AddSP dependent on the former memory */
917 add_irn_dep(subsp, get_Free_mem(free));
920 exchange(free, sync);
927 * Check if a node is somehow data dependent on another one.
928 * both nodes must be in the same basic block.
929 * @param n1 The first node.
930 * @param n2 The second node.
931 * @return 1, if n1 is data dependent (transitively) on n2, 0 if not.
933 static int dependent_on(ir_node *n1, ir_node *n2)
935 assert(get_nodes_block(n1) == get_nodes_block(n2));
937 return heights_reachable_in_block(ir_heights, n1, n2);
940 static int cmp_call_dependency(const void *c1, const void *c2)
942 ir_node *n1 = *(ir_node **) c1;
943 ir_node *n2 = *(ir_node **) c2;
947 Classical qsort() comparison function behavior:
948 0 if both elements are equal
949 1 if second is "smaller" that first
950 -1 if first is "smaller" that second
952 if (dependent_on(n1, n2))
955 if (dependent_on(n2, n1))
958 /* The nodes have no depth order, but we need a total order because qsort()
961 * Additionally, we need to respect transitive dependencies. Consider a
962 * Call a depending on Call b and an independent Call c.
963 * We MUST NOT order c > a and b > c. */
964 h1 = get_irn_height(ir_heights, n1);
965 h2 = get_irn_height(ir_heights, n2);
966 if (h1 < h2) return -1;
967 if (h1 > h2) return 1;
968 /* Same height, so use a random (but stable) order */
969 return get_irn_idx(n1) - get_irn_idx(n2);
973 * Walker: links all Call/Alloc/Free nodes to the Block they are contained.
975 static void link_ops_in_block_walker(ir_node *irn, void *data)
977 be_abi_irg_t *env = (be_abi_irg_t*)data;
978 unsigned code = get_irn_opcode(irn);
980 if (code == iro_Call ||
981 (code == iro_Alloc && get_Alloc_where(irn) == stack_alloc) ||
982 (code == iro_Free && get_Free_where(irn) == stack_alloc)) {
983 ir_node *bl = get_nodes_block(irn);
984 void *save = get_irn_link(bl);
986 set_irn_link(irn, save);
987 set_irn_link(bl, irn);
990 if (code == iro_Builtin && get_Builtin_kind(irn) == ir_bk_return_address) {
991 ir_node *param = get_Builtin_param(irn, 0);
992 ir_tarval *tv = get_Const_tarval(param);
993 unsigned long value = get_tarval_long(tv);
994 /* use ebp, so the climbframe algo works... */
996 env->call->flags.bits.try_omit_fp = 0;
1003 * Process all Call/Alloc/Free nodes inside a basic block.
1004 * Note that the link field of the block must contain a linked list of all
1005 * nodes inside the Block. We first order this list according to data dependency
1006 * and that connect the nodes together.
1008 static void process_ops_in_block(ir_node *bl, void *data)
1010 be_abi_irg_t *env = (be_abi_irg_t*)data;
1011 ir_node *curr_sp = env->init_sp;
1018 for (irn = (ir_node*)get_irn_link(bl); irn != NULL;
1019 irn = (ir_node*)get_irn_link(irn)) {
1023 nodes = ALLOCAN(ir_node*, n_nodes);
1024 for (irn = (ir_node*)get_irn_link(bl), n = 0; irn != NULL;
1025 irn = (ir_node*)get_irn_link(irn), ++n) {
1029 /* If there were call nodes in the block. */
1034 /* order the call nodes according to data dependency */
1035 qsort(nodes, n_nodes, sizeof(nodes[0]), cmp_call_dependency);
1037 for (i = n_nodes - 1; i >= 0; --i) {
1038 ir_node *irn = nodes[i];
1040 DBG((dbg, LEVEL_3, "\tprocessing call %+F\n", irn));
1041 switch (get_irn_opcode(irn)) {
1044 /* The stack pointer will be modified due to a call. */
1045 env->call->flags.bits.try_omit_fp = 0;
1047 curr_sp = adjust_call(env, irn, curr_sp);
1050 if (get_Alloc_where(irn) == stack_alloc)
1051 curr_sp = adjust_alloc(env, irn, curr_sp);
1054 if (get_Free_where(irn) == stack_alloc)
1055 curr_sp = adjust_free(env, irn, curr_sp);
1058 panic("invalid call");
1062 /* Keep the last stack state in the block by tying it to Keep node,
1063 * the proj from calls is already kept */
1064 if (curr_sp != env->init_sp &&
1065 !(is_Proj(curr_sp) && be_is_Call(get_Proj_pred(curr_sp)))) {
1067 keep = be_new_Keep(bl, 1, nodes);
1068 pmap_insert(env->keep_map, bl, keep);
1072 set_irn_link(bl, curr_sp);
1076 * Adjust all call nodes in the graph to the ABI conventions.
1078 static void process_calls(ir_graph *const irg, be_abi_irg_t *const abi)
1080 irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, abi);
1082 ir_heights = heights_new(irg);
1083 irg_block_walk_graph(irg, NULL, process_ops_in_block, abi);
1084 heights_free(ir_heights);
1088 * Computes the stack argument layout type.
1089 * Changes a possibly allocated value param type by moving
1090 * entities to the stack layout type.
1092 * @param call the current call ABI
1093 * @param method_type the method type
1095 * @return the stack argument layout type
1097 static ir_type *compute_arg_type(ir_graph *irg, be_abi_call_t *call,
1098 ir_type *method_type)
1100 struct obstack *obst = be_get_be_obst(irg);
1101 ir_type *frame_type = get_irg_frame_type(irg);
1102 size_t n_params = get_method_n_params(method_type);
1103 size_t n_frame_members = get_compound_n_members(frame_type);
1104 ir_entity *va_start_entity = NULL;
1110 ir_entity **map = OALLOCNZ(obst, ir_entity*, n_params);
1111 res = new_type_struct(new_id_from_chars("arg_type", 8));
1113 /* collect existing entities for value_param_types */
1114 for (f = n_frame_members; f > 0; ) {
1115 ir_entity *entity = get_compound_member(frame_type, --f);
1118 set_entity_link(entity, NULL);
1119 if (!is_parameter_entity(entity))
1121 num = get_entity_parameter_number(entity);
1122 if (num == IR_VA_START_PARAMETER_NUMBER) {
1123 /* move entity to new arg_type */
1124 set_entity_owner(entity, res);
1125 va_start_entity = entity;
1128 assert(num < n_params);
1129 if (map[num] != NULL)
1130 panic("multiple entities for parameter %u in %+F found", f, irg);
1132 if (num != n_params && !get_call_arg(call, 0, num, 1)->on_stack) {
1133 /* don't move this entity */
1138 /* move entity to new arg_type */
1139 set_entity_owner(entity, res);
1142 for (i = 0; i < n_params; ++i) {
1143 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1144 ir_type *param_type = get_method_param_type(method_type, i);
1147 if (!arg->on_stack) {
1151 if (entity == NULL) {
1152 /* create a new entity */
1153 entity = new_parameter_entity(res, i, param_type);
1155 ofs += arg->space_before;
1156 ofs = round_up2(ofs, arg->alignment);
1157 set_entity_offset(entity, ofs);
1158 ofs += arg->space_after;
1159 ofs += get_type_size_bytes(param_type);
1160 arg->stack_ent = entity;
1162 if (va_start_entity != NULL) {
1163 set_entity_offset(va_start_entity, ofs);
1165 set_type_size_bytes(res, ofs);
1166 set_type_state(res, layout_fixed);
1172 const arch_register_t *reg;
1176 static int cmp_regs(const void *a, const void *b)
1178 const reg_node_map_t *p = (const reg_node_map_t*)a;
1179 const reg_node_map_t *q = (const reg_node_map_t*)b;
1181 if (p->reg->reg_class == q->reg->reg_class)
1182 return p->reg->index - q->reg->index;
1184 return p->reg->reg_class < q->reg->reg_class ? -1 : +1;
1187 static void reg_map_to_arr(reg_node_map_t *res, pmap *reg_map)
1190 size_t n = pmap_count(reg_map);
1193 foreach_pmap(reg_map, ent) {
1194 res[i].reg = (const arch_register_t*)ent->key;
1195 res[i].irn = (ir_node*)ent->value;
1199 qsort(res, n, sizeof(res[0]), cmp_regs);
1203 * Creates a be_Return for a Return node.
1205 * @param @env the abi environment
1206 * @param irn the Return node or NULL if there was none
1207 * @param bl the block where the be_Retun should be placed
1208 * @param mem the current memory
1209 * @param n_res number of return results
1211 static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl,
1212 ir_node *mem, int n_res)
1214 be_abi_call_t *call = env->call;
1215 ir_graph *irg = get_Block_irg(bl);
1216 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1218 pmap *reg_map = pmap_create();
1219 ir_node *keep = pmap_get(ir_node, env->keep_map, bl);
1226 const arch_register_t **regs;
1230 get the valid stack node in this block.
1231 If we had a call in that block there is a Keep constructed by process_calls()
1232 which points to the last stack modification in that block. we'll use
1233 it then. Else we use the stack from the start block and let
1234 the ssa construction fix the usage.
1236 stack = be_abi_reg_map_get(env->regs, arch_env->sp);
1238 stack = get_irn_n(keep, 0);
1240 remove_End_keepalive(get_irg_end(irg), keep);
1243 /* Insert results for Return into the register map. */
1244 for (i = 0; i < n_res; ++i) {
1245 ir_node *res = get_Return_res(irn, i);
1246 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1247 assert(arg->in_reg && "return value must be passed in register");
1248 pmap_insert(reg_map, (void *) arg->reg, res);
1251 /* Add uses of the callee save registers. */
1252 foreach_pmap(env->regs, ent) {
1253 const arch_register_t *reg = (const arch_register_t*)ent->key;
1254 if ((reg->type & arch_register_type_ignore) || arch_register_is_callee_save(arch_env, reg))
1255 pmap_insert(reg_map, ent->key, ent->value);
1258 be_abi_reg_map_set(reg_map, arch_env->sp, stack);
1261 Maximum size of the in array for Return nodes is
1262 return args + callee save/ignore registers + memory + stack pointer
1264 in_max = pmap_count(reg_map) + n_res + 2;
1266 in = ALLOCAN(ir_node*, in_max);
1267 regs = ALLOCAN(arch_register_t const*, in_max);
1270 in[1] = be_abi_reg_map_get(reg_map, arch_env->sp);
1272 regs[1] = arch_env->sp;
1275 /* clear SP entry, since it has already been grown. */
1276 pmap_insert(reg_map, (void *) arch_env->sp, NULL);
1277 for (i = 0; i < n_res; ++i) {
1278 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1280 in[n] = be_abi_reg_map_get(reg_map, arg->reg);
1281 regs[n++] = arg->reg;
1283 /* Clear the map entry to mark the register as processed. */
1284 be_abi_reg_map_set(reg_map, arg->reg, NULL);
1287 /* grow the rest of the stuff. */
1288 foreach_pmap(reg_map, ent) {
1290 in[n] = (ir_node*)ent->value;
1291 regs[n++] = (const arch_register_t*)ent->key;
1295 /* The in array for the new back end return is now ready. */
1297 dbgi = get_irn_dbg_info(irn);
1301 /* we have to pop the shadow parameter in in case of struct returns */
1303 ret = be_new_Return(dbgi, irg, bl, n_res, pop, n, in);
1305 /* Set the register classes of the return's parameter accordingly. */
1306 for (i = 0; i < n; ++i) {
1307 if (regs[i] == NULL)
1310 be_set_constr_single_reg_in(ret, i, regs[i], arch_register_req_type_none);
1313 /* Free the space of the Epilog's in array and the register <-> proj map. */
1314 pmap_destroy(reg_map);
1319 typedef struct lower_frame_sels_env_t {
1320 ir_node *frame; /**< the current frame */
1321 const arch_register_class_t *sp_class; /**< register class of the stack pointer */
1322 } lower_frame_sels_env_t;
1325 * Walker: Replaces Sels of frame type and
1326 * value param type entities by FrameAddress.
1327 * Links all used entities.
1329 static void lower_frame_sels_walker(ir_node *irn, void *data)
1331 lower_frame_sels_env_t *ctx = (lower_frame_sels_env_t*)data;
1334 ir_node *ptr = get_Sel_ptr(irn);
1336 if (ptr == ctx->frame) {
1337 ir_entity *ent = get_Sel_entity(irn);
1338 ir_node *bl = get_nodes_block(irn);
1341 nw = be_new_FrameAddr(ctx->sp_class, bl, ctx->frame, ent);
1348 * The start block has no jump, instead it has an initial exec Proj.
1349 * The backend wants to handle all blocks the same way, so we replace
1350 * the out cfg edge with a real jump.
1352 static void fix_start_block(ir_graph *irg)
1354 ir_node *initial_X = get_irg_initial_exec(irg);
1355 ir_node *start_block = get_irg_start_block(irg);
1356 ir_node *jmp = new_r_Jmp(start_block);
1358 assert(is_Proj(initial_X));
1359 exchange(initial_X, jmp);
1360 set_irg_initial_exec(irg, new_r_Bad(irg, mode_X));
1362 /* merge start block with successor if possible */
1364 foreach_out_edge(jmp, edge) {
1365 ir_node *succ = get_edge_src_irn(edge);
1366 if (!is_Block(succ))
1369 if (get_irn_arity(succ) == 1) {
1370 exchange(succ, start_block);
1378 * Modify the irg itself and the frame type.
1380 static void modify_irg(ir_graph *const irg, be_abi_irg_t *const env)
1382 be_abi_call_t *call = env->call;
1383 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1384 const arch_register_t *sp = arch_env->sp;
1385 ir_type *method_type = get_entity_type(get_irg_entity(irg));
1386 be_irg_t *birg = be_birg_from_irg(irg);
1387 struct obstack *obst = be_get_be_obst(irg);
1388 be_stack_layout_t *stack_layout = be_get_irg_stack_layout(irg);
1391 ir_node *new_mem_proj;
1397 unsigned frame_size;
1400 const arch_register_t *fp_reg;
1401 ir_node *frame_pointer;
1405 ir_type *arg_type, *bet_type;
1406 lower_frame_sels_env_t ctx;
1408 DBG((dbg, LEVEL_1, "introducing abi on %+F\n", irg));
1410 old_mem = get_irg_initial_mem(irg);
1412 irp_reserve_resources(irp, IRP_RESOURCE_ENTITY_LINK);
1414 arg_type = compute_arg_type(irg, call, method_type);
1416 /* Convert the Sel nodes in the irg to frame addr nodes: */
1417 ctx.frame = get_irg_frame(irg);
1418 ctx.sp_class = arch_env->sp->reg_class;
1420 ir_type *const frame_tp = get_irg_frame_type(irg);
1421 /* layout the stackframe now */
1422 if (get_type_state(frame_tp) == layout_undefined) {
1423 default_layout_compound_type(frame_tp);
1426 /* align stackframe to 4 byte */
1427 frame_size = get_type_size_bytes(frame_tp);
1428 if (frame_size % 4 != 0) {
1429 set_type_size_bytes(frame_tp, frame_size + 4 - (frame_size % 4));
1432 env->regs = pmap_create();
1434 n_params = get_method_n_params(method_type);
1435 args = OALLOCNZ(obst, ir_node*, n_params);
1437 be_add_parameter_entity_stores(irg);
1439 irg_walk_graph(irg, lower_frame_sels_walker, NULL, &ctx);
1441 irp_free_resources(irp, IRP_RESOURCE_ENTITY_LINK);
1443 /* Fill the argument vector */
1444 arg_tuple = get_irg_args(irg);
1445 foreach_out_edge(arg_tuple, edge) {
1446 ir_node *irn = get_edge_src_irn(edge);
1447 if (! is_Anchor(irn)) {
1448 int nr = get_Proj_proj(irn);
1450 DBG((dbg, LEVEL_2, "\treading arg: %d -> %+F\n", nr, irn));
1454 stack_layout->sp_relative = call->flags.bits.try_omit_fp;
1455 bet_type = call->cb->get_between_type(irg);
1456 stack_frame_init(stack_layout, arg_type, bet_type,
1457 get_irg_frame_type(irg));
1459 /* Count the register params and add them to the number of Projs for the RegParams node */
1460 for (i = 0; i < n_params; ++i) {
1461 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1462 if (arg->in_reg && args[i]) {
1463 assert(arg->reg != sp && "cannot use stack pointer as parameter register");
1464 assert(i == get_Proj_proj(args[i]));
1466 /* For now, associate the register with the old Proj from Start representing that argument. */
1467 pmap_insert(env->regs, (void *) arg->reg, args[i]);
1468 DBG((dbg, LEVEL_2, "\targ #%d -> reg %s\n", i, arg->reg->name));
1472 /* Collect all callee-save registers */
1473 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
1474 const arch_register_class_t *cls = &arch_env->register_classes[i];
1475 for (j = 0; j < cls->n_regs; ++j) {
1476 const arch_register_t *reg = &cls->regs[j];
1477 if ((reg->type & arch_register_type_state) || arch_register_is_callee_save(arch_env, reg)) {
1478 pmap_insert(env->regs, (void *) reg, NULL);
1483 fp_reg = call->flags.bits.try_omit_fp ? arch_env->sp : arch_env->bp;
1484 rbitset_clear(birg->allocatable_regs, fp_reg->global_index);
1486 /* handle start block here (place a jump in the block) */
1487 fix_start_block(irg);
1489 pmap_insert(env->regs, (void *) sp, NULL);
1490 pmap_insert(env->regs, (void *) arch_env->bp, NULL);
1491 start_bl = get_irg_start_block(irg);
1492 ir_node *const start = be_new_Start(NULL, start_bl, pmap_count(env->regs) + 1);
1493 set_irg_start(irg, start);
1496 * make proj nodes for the callee save registers.
1497 * memorize them, since Return nodes get those as inputs.
1499 * Note, that if a register corresponds to an argument, the regs map
1500 * contains the old Proj from start for that argument.
1502 rm = ALLOCAN(reg_node_map_t, pmap_count(env->regs));
1503 reg_map_to_arr(rm, env->regs);
1504 for (i = 0, n = pmap_count(env->regs); i < n; ++i) {
1505 const arch_register_t *reg = rm[i].reg;
1506 ir_mode *mode = reg->reg_class->mode;
1508 arch_register_req_type_t add_type = arch_register_req_type_none;
1512 add_type |= arch_register_req_type_produces_sp;
1513 if (!rbitset_is_set(birg->allocatable_regs, reg->global_index)) {
1514 add_type |= arch_register_req_type_ignore;
1518 proj = new_r_Proj(start, mode, nr + 1);
1519 pmap_insert(env->regs, (void *) reg, proj);
1520 be_set_constr_single_reg_out(start, nr + 1, reg, add_type);
1521 arch_set_irn_register(proj, reg);
1523 DBG((dbg, LEVEL_2, "\tregister save proj #%d -> reg %s\n", nr, reg->name));
1526 /* create a new initial memory proj */
1527 assert(is_Proj(old_mem));
1528 arch_set_irn_register_req_out(start, 0, arch_no_register_req);
1529 new_mem_proj = new_r_Proj(start, mode_M, 0);
1531 set_irg_initial_mem(irg, mem);
1533 env->init_sp = be_abi_reg_map_get(env->regs, sp);
1535 /* set new frame_pointer */
1536 frame_pointer = be_abi_reg_map_get(env->regs, fp_reg);
1537 set_irg_frame(irg, frame_pointer);
1539 /* rewire old mem users to new mem */
1540 exchange(old_mem, mem);
1542 /* keep the mem (for functions with an endless loop = no return) */
1545 set_irg_initial_mem(irg, mem);
1547 /* Now, introduce stack param nodes for all parameters passed on the stack */
1548 for (i = 0; i < n_params; ++i) {
1549 ir_node *arg_proj = args[i];
1550 ir_node *repl = NULL;
1552 if (arg_proj != NULL) {
1553 be_abi_call_arg_t *arg;
1554 ir_type *param_type;
1555 int nr = get_Proj_proj(arg_proj);
1558 nr = MIN(nr, n_params);
1559 arg = get_call_arg(call, 0, nr, 1);
1560 param_type = get_method_param_type(method_type, nr);
1563 repl = pmap_get(ir_node, env->regs, arg->reg);
1564 } else if (arg->on_stack) {
1565 ir_node *addr = be_new_FrameAddr(sp->reg_class, start_bl, frame_pointer, arg->stack_ent);
1567 /* For atomic parameters which are actually used, we create a Load node. */
1568 if (is_atomic_type(param_type) && get_irn_n_edges(args[i]) > 0) {
1569 ir_mode *mode = get_type_mode(param_type);
1570 ir_mode *load_mode = arg->load_mode;
1571 ir_node *nomem = get_irg_no_mem(irg);
1573 ir_node *load = new_r_Load(start_bl, nomem, addr, load_mode, cons_floats);
1574 repl = new_r_Proj(load, load_mode, pn_Load_res);
1576 if (mode != load_mode) {
1577 repl = new_r_Conv(start_bl, repl, mode);
1580 /* The stack parameter is not primitive (it is a struct or array),
1581 * we thus will create a node representing the parameter's address
1587 assert(repl != NULL);
1589 /* Beware: the mode of the register parameters is always the mode of the register class
1590 which may be wrong. Add Conv's then. */
1591 mode = get_irn_mode(args[i]);
1592 if (mode != get_irn_mode(repl)) {
1593 repl = new_r_Conv(get_nodes_block(repl), repl, mode);
1595 exchange(args[i], repl);
1599 /* the arg proj is not needed anymore now and should be only used by the anchor */
1600 assert(get_irn_n_edges(arg_tuple) == 1);
1601 kill_node(arg_tuple);
1602 set_irg_args(irg, new_r_Bad(irg, mode_T));
1604 /* All Return nodes hang on the End node, so look for them there. */
1605 end = get_irg_end_block(irg);
1606 for (i = 0, n = get_Block_n_cfgpreds(end); i < n; ++i) {
1607 ir_node *irn = get_Block_cfgpred(end, i);
1609 if (is_Return(irn)) {
1610 ir_node *blk = get_nodes_block(irn);
1611 ir_node *mem = get_Return_mem(irn);
1612 ir_node *ret = create_be_return(env, irn, blk, mem, get_Return_n_ress(irn));
1617 /* if we have endless loops here, n might be <= 0. Do NOT create a be_Return then,
1618 the code is dead and will never be executed. */
1621 /** Fix the state inputs of calls that still hang on unknowns */
1622 static void fix_call_state_inputs(ir_graph *const irg, be_abi_irg_t *const env)
1624 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1626 arch_register_t **stateregs = NEW_ARR_F(arch_register_t*, 0);
1628 /* Collect caller save registers */
1629 n = arch_env->n_register_classes;
1630 for (i = 0; i < n; ++i) {
1632 const arch_register_class_t *cls = &arch_env->register_classes[i];
1633 for (j = 0; j < cls->n_regs; ++j) {
1634 const arch_register_t *reg = arch_register_for_index(cls, j);
1635 if (reg->type & arch_register_type_state) {
1636 ARR_APP1(arch_register_t*, stateregs, (arch_register_t *)reg);
1641 n = ARR_LEN(env->calls);
1642 n_states = ARR_LEN(stateregs);
1643 for (i = 0; i < n; ++i) {
1645 ir_node *call = env->calls[i];
1647 arity = get_irn_arity(call);
1649 /* the state reg inputs are the last n inputs of the calls */
1650 for (s = 0; s < n_states; ++s) {
1651 int inp = arity - n_states + s;
1652 const arch_register_t *reg = stateregs[s];
1653 ir_node *regnode = be_abi_reg_map_get(env->regs, reg);
1655 set_irn_n(call, inp, regnode);
1659 DEL_ARR_F(stateregs);
1663 * Create a trampoline entity for the given method.
1665 static ir_entity *create_trampoline(be_main_env_t *be, ir_entity *method)
1667 ir_type *type = get_entity_type(method);
1668 ident *old_id = get_entity_ld_ident(method);
1669 ident *id = id_mangle3("", old_id, "$stub");
1670 ir_type *parent = be->pic_trampolines_type;
1671 ir_entity *ent = new_entity(parent, old_id, type);
1672 set_entity_ld_ident(ent, id);
1673 set_entity_visibility(ent, ir_visibility_private);
1679 * Returns the trampoline entity for the given method.
1681 static ir_entity *get_trampoline(be_main_env_t *env, ir_entity *method)
1683 ir_entity *result = pmap_get(ir_entity, env->ent_trampoline_map, method);
1684 if (result == NULL) {
1685 result = create_trampoline(env, method);
1686 pmap_insert(env->ent_trampoline_map, method, result);
1692 static ir_entity *create_pic_symbol(be_main_env_t *be, ir_entity *entity)
1694 ident *old_id = get_entity_ld_ident(entity);
1695 ident *id = id_mangle3("", old_id, "$non_lazy_ptr");
1696 ir_type *e_type = get_entity_type(entity);
1697 ir_type *type = new_type_pointer(e_type);
1698 ir_type *parent = be->pic_symbols_type;
1699 ir_entity *ent = new_entity(parent, old_id, type);
1700 set_entity_ld_ident(ent, id);
1701 set_entity_visibility(ent, ir_visibility_private);
1706 static ir_entity *get_pic_symbol(be_main_env_t *env, ir_entity *entity)
1708 ir_entity *result = pmap_get(ir_entity, env->ent_pic_symbol_map, entity);
1709 if (result == NULL) {
1710 result = create_pic_symbol(env, entity);
1711 pmap_insert(env->ent_pic_symbol_map, entity, result);
1720 * Returns non-zero if a given entity can be accessed using a relative address.
1722 static int can_address_relative(ir_entity *entity)
1724 return entity_has_definition(entity) && !(get_entity_linkage(entity) & IR_LINKAGE_MERGE);
1727 static ir_node *get_pic_base(ir_graph *irg)
1729 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1730 if (arch_env->impl->get_pic_base == NULL)
1732 return arch_env->impl->get_pic_base(irg);
1735 /** patches SymConsts to work in position independent code */
1736 static void fix_pic_symconsts(ir_node *node, void *data)
1738 ir_graph *irg = get_irn_irg(node);
1739 be_main_env_t *be = be_get_irg_main_env(irg);
1749 arity = get_irn_arity(node);
1750 for (i = 0; i < arity; ++i) {
1752 ir_node *pred = get_irn_n(node, i);
1754 ir_entity *pic_symbol;
1755 ir_node *pic_symconst;
1757 if (!is_SymConst(pred))
1760 entity = get_SymConst_entity(pred);
1761 block = get_nodes_block(pred);
1763 /* calls can jump to relative addresses, so we can directly jump to
1764 the (relatively) known call address or the trampoline */
1765 if (i == 1 && is_Call(node)) {
1766 ir_entity *trampoline;
1767 ir_node *trampoline_const;
1769 if (can_address_relative(entity))
1772 dbgi = get_irn_dbg_info(pred);
1773 trampoline = get_trampoline(be, entity);
1774 trampoline_const = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
1776 set_irn_n(node, i, trampoline_const);
1780 /* everything else is accessed relative to EIP */
1781 mode = get_irn_mode(pred);
1782 pic_base = get_pic_base(irg);
1784 /* all ok now for locally constructed stuff */
1785 if (can_address_relative(entity)) {
1786 ir_node *add = new_r_Add(block, pic_base, pred, mode);
1788 /* make sure the walker doesn't visit this add again */
1789 mark_irn_visited(add);
1790 set_irn_n(node, i, add);
1794 /* get entry from pic symbol segment */
1795 dbgi = get_irn_dbg_info(pred);
1796 pic_symbol = get_pic_symbol(be, entity);
1797 pic_symconst = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
1799 add = new_r_Add(block, pic_base, pic_symconst, mode);
1800 mark_irn_visited(add);
1802 /* we need an extra indirection for global data outside our current
1803 module. The loads are always safe and can therefore float
1804 and need no memory input */
1805 load = new_r_Load(block, get_irg_no_mem(irg), add, mode, cons_floats);
1806 load_res = new_r_Proj(load, mode, pn_Load_res);
1808 set_irn_n(node, i, load_res);
1812 void be_abi_introduce(ir_graph *irg)
1814 ir_node *old_frame = get_irg_frame(irg);
1815 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1816 ir_entity *entity = get_irg_entity(irg);
1817 ir_type *method_type = get_entity_type(entity);
1818 be_irg_t *birg = be_birg_from_irg(irg);
1819 struct obstack *obst = &birg->obst;
1820 ir_node *dummy = new_r_Dummy(irg,
1821 arch_env->sp->reg_class->mode);
1824 /* determine allocatable registers */
1825 assert(birg->allocatable_regs == NULL);
1826 birg->allocatable_regs = rbitset_obstack_alloc(obst, arch_env->n_registers);
1827 for (r = 0; r < arch_env->n_registers; ++r) {
1828 const arch_register_t *reg = &arch_env->registers[r];
1829 if ( !(reg->type & arch_register_type_ignore)) {
1830 rbitset_set(birg->allocatable_regs, r);
1834 /* Break here if backend provides a custom API. */
1836 be_omit_fp = be_options.omit_fp;
1839 env.keep_map = pmap_create();
1840 env.call = be_abi_call_new(arch_env->sp->reg_class);
1841 arch_env_get_call_abi(arch_env, method_type, env.call);
1843 env.init_sp = dummy;
1844 env.calls = NEW_ARR_F(ir_node*, 0);
1848 if (be_options.pic) {
1849 irg_walk_graph(irg, fix_pic_symconsts, NULL, NULL);
1852 /* Lower all call nodes in the IRG. */
1853 process_calls(irg, &env);
1855 /* Process the IRG */
1856 modify_irg(irg, &env);
1858 /* fix call inputs for state registers */
1859 fix_call_state_inputs(irg, &env);
1861 be_abi_call_free(env.call);
1863 /* We don't need the keep map anymore. */
1864 pmap_destroy(env.keep_map);
1866 /* calls array is not needed anymore */
1867 DEL_ARR_F(env.calls);
1869 /* reroute the stack origin of the calls to the true stack origin. */
1870 exchange(dummy, env.init_sp);
1871 exchange(old_frame, get_irg_frame(irg));
1873 pmap_destroy(env.regs);
1876 void be_put_allocatable_regs(const ir_graph *irg,
1877 const arch_register_class_t *cls, bitset_t *bs)
1879 be_irg_t *birg = be_birg_from_irg(irg);
1880 unsigned *allocatable_regs = birg->allocatable_regs;
1883 assert(bitset_size(bs) == cls->n_regs);
1884 bitset_clear_all(bs);
1885 for (i = 0; i < cls->n_regs; ++i) {
1886 const arch_register_t *reg = &cls->regs[i];
1887 if (rbitset_is_set(allocatable_regs, reg->global_index))
1892 unsigned be_get_n_allocatable_regs(const ir_graph *irg,
1893 const arch_register_class_t *cls)
1895 bitset_t *bs = bitset_alloca(cls->n_regs);
1896 be_put_allocatable_regs(irg, cls, bs);
1897 return bitset_popcount(bs);
1900 void be_set_allocatable_regs(const ir_graph *irg,
1901 const arch_register_class_t *cls,
1902 unsigned *raw_bitset)
1904 be_irg_t *birg = be_birg_from_irg(irg);
1905 unsigned *allocatable_regs = birg->allocatable_regs;
1908 rbitset_clear_all(raw_bitset, cls->n_regs);
1909 for (i = 0; i < cls->n_regs; ++i) {
1910 const arch_register_t *reg = &cls->regs[i];
1911 if (rbitset_is_set(allocatable_regs, reg->global_index))
1912 rbitset_set(raw_bitset, i);
1916 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_abi)
1917 void be_init_abi(void)
1919 FIRM_DBG_REGISTER(dbg, "firm.be.abi");