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 *irg)
1080 be_abi_irg_t *abi = be_get_irg_abi(irg);
1082 irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, abi);
1084 ir_heights = heights_new(irg);
1085 irg_block_walk_graph(irg, NULL, process_ops_in_block, abi);
1086 heights_free(ir_heights);
1090 * Computes the stack argument layout type.
1091 * Changes a possibly allocated value param type by moving
1092 * entities to the stack layout type.
1094 * @param call the current call ABI
1095 * @param method_type the method type
1097 * @return the stack argument layout type
1099 static ir_type *compute_arg_type(ir_graph *irg, be_abi_call_t *call,
1100 ir_type *method_type)
1102 struct obstack *obst = be_get_be_obst(irg);
1103 ir_type *frame_type = get_irg_frame_type(irg);
1104 size_t n_params = get_method_n_params(method_type);
1105 size_t n_frame_members = get_compound_n_members(frame_type);
1106 ir_entity *va_start_entity = NULL;
1112 ir_entity **map = OALLOCNZ(obst, ir_entity*, n_params);
1113 res = new_type_struct(new_id_from_chars("arg_type", 8));
1115 /* collect existing entities for value_param_types */
1116 for (f = n_frame_members; f > 0; ) {
1117 ir_entity *entity = get_compound_member(frame_type, --f);
1120 set_entity_link(entity, NULL);
1121 if (!is_parameter_entity(entity))
1123 num = get_entity_parameter_number(entity);
1124 if (num == IR_VA_START_PARAMETER_NUMBER) {
1125 /* move entity to new arg_type */
1126 set_entity_owner(entity, res);
1127 va_start_entity = entity;
1130 assert(num < n_params);
1131 if (map[num] != NULL)
1132 panic("multiple entities for parameter %u in %+F found", f, irg);
1134 if (num != n_params && !get_call_arg(call, 0, num, 1)->on_stack) {
1135 /* don't move this entity */
1140 /* move entity to new arg_type */
1141 set_entity_owner(entity, res);
1144 for (i = 0; i < n_params; ++i) {
1145 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1146 ir_type *param_type = get_method_param_type(method_type, i);
1149 if (!arg->on_stack) {
1153 if (entity == NULL) {
1154 /* create a new entity */
1155 entity = new_parameter_entity(res, i, param_type);
1157 ofs += arg->space_before;
1158 ofs = round_up2(ofs, arg->alignment);
1159 set_entity_offset(entity, ofs);
1160 ofs += arg->space_after;
1161 ofs += get_type_size_bytes(param_type);
1162 arg->stack_ent = entity;
1164 if (va_start_entity != NULL) {
1165 set_entity_offset(va_start_entity, ofs);
1167 set_type_size_bytes(res, ofs);
1168 set_type_state(res, layout_fixed);
1174 const arch_register_t *reg;
1178 static int cmp_regs(const void *a, const void *b)
1180 const reg_node_map_t *p = (const reg_node_map_t*)a;
1181 const reg_node_map_t *q = (const reg_node_map_t*)b;
1183 if (p->reg->reg_class == q->reg->reg_class)
1184 return p->reg->index - q->reg->index;
1186 return p->reg->reg_class < q->reg->reg_class ? -1 : +1;
1189 static void reg_map_to_arr(reg_node_map_t *res, pmap *reg_map)
1192 size_t n = pmap_count(reg_map);
1195 foreach_pmap(reg_map, ent) {
1196 res[i].reg = (const arch_register_t*)ent->key;
1197 res[i].irn = (ir_node*)ent->value;
1201 qsort(res, n, sizeof(res[0]), cmp_regs);
1205 * Creates a be_Return for a Return node.
1207 * @param @env the abi environment
1208 * @param irn the Return node or NULL if there was none
1209 * @param bl the block where the be_Retun should be placed
1210 * @param mem the current memory
1211 * @param n_res number of return results
1213 static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl,
1214 ir_node *mem, int n_res)
1216 be_abi_call_t *call = env->call;
1217 ir_graph *irg = get_Block_irg(bl);
1218 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1220 pmap *reg_map = pmap_create();
1221 ir_node *keep = pmap_get(ir_node, env->keep_map, bl);
1228 const arch_register_t **regs;
1232 get the valid stack node in this block.
1233 If we had a call in that block there is a Keep constructed by process_calls()
1234 which points to the last stack modification in that block. we'll use
1235 it then. Else we use the stack from the start block and let
1236 the ssa construction fix the usage.
1238 stack = be_abi_reg_map_get(env->regs, arch_env->sp);
1240 stack = get_irn_n(keep, 0);
1242 remove_End_keepalive(get_irg_end(irg), keep);
1245 /* Insert results for Return into the register map. */
1246 for (i = 0; i < n_res; ++i) {
1247 ir_node *res = get_Return_res(irn, i);
1248 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1249 assert(arg->in_reg && "return value must be passed in register");
1250 pmap_insert(reg_map, (void *) arg->reg, res);
1253 /* Add uses of the callee save registers. */
1254 foreach_pmap(env->regs, ent) {
1255 const arch_register_t *reg = (const arch_register_t*)ent->key;
1256 if ((reg->type & arch_register_type_ignore) || arch_register_is_callee_save(arch_env, reg))
1257 pmap_insert(reg_map, ent->key, ent->value);
1260 be_abi_reg_map_set(reg_map, arch_env->sp, stack);
1263 Maximum size of the in array for Return nodes is
1264 return args + callee save/ignore registers + memory + stack pointer
1266 in_max = pmap_count(reg_map) + n_res + 2;
1268 in = ALLOCAN(ir_node*, in_max);
1269 regs = ALLOCAN(arch_register_t const*, in_max);
1272 in[1] = be_abi_reg_map_get(reg_map, arch_env->sp);
1274 regs[1] = arch_env->sp;
1277 /* clear SP entry, since it has already been grown. */
1278 pmap_insert(reg_map, (void *) arch_env->sp, NULL);
1279 for (i = 0; i < n_res; ++i) {
1280 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1282 in[n] = be_abi_reg_map_get(reg_map, arg->reg);
1283 regs[n++] = arg->reg;
1285 /* Clear the map entry to mark the register as processed. */
1286 be_abi_reg_map_set(reg_map, arg->reg, NULL);
1289 /* grow the rest of the stuff. */
1290 foreach_pmap(reg_map, ent) {
1292 in[n] = (ir_node*)ent->value;
1293 regs[n++] = (const arch_register_t*)ent->key;
1297 /* The in array for the new back end return is now ready. */
1299 dbgi = get_irn_dbg_info(irn);
1303 /* we have to pop the shadow parameter in in case of struct returns */
1305 ret = be_new_Return(dbgi, irg, bl, n_res, pop, n, in);
1307 /* Set the register classes of the return's parameter accordingly. */
1308 for (i = 0; i < n; ++i) {
1309 if (regs[i] == NULL)
1312 be_set_constr_single_reg_in(ret, i, regs[i], arch_register_req_type_none);
1315 /* Free the space of the Epilog's in array and the register <-> proj map. */
1316 pmap_destroy(reg_map);
1321 typedef struct lower_frame_sels_env_t {
1322 ir_node *frame; /**< the current frame */
1323 const arch_register_class_t *sp_class; /**< register class of the stack pointer */
1324 const arch_register_class_t *link_class; /**< register class of the link pointer */
1325 ir_type *frame_tp; /**< the frame type */
1326 int static_link_pos; /**< argument number of the hidden static link */
1327 } lower_frame_sels_env_t;
1330 * Walker: Replaces Sels of frame type and
1331 * value param type entities by FrameAddress.
1332 * Links all used entities.
1334 static void lower_frame_sels_walker(ir_node *irn, void *data)
1336 lower_frame_sels_env_t *ctx = (lower_frame_sels_env_t*)data;
1339 ir_node *ptr = get_Sel_ptr(irn);
1341 if (ptr == ctx->frame) {
1342 ir_entity *ent = get_Sel_entity(irn);
1343 ir_node *bl = get_nodes_block(irn);
1346 nw = be_new_FrameAddr(ctx->sp_class, bl, ctx->frame, ent);
1353 * The start block has no jump, instead it has an initial exec Proj.
1354 * The backend wants to handle all blocks the same way, so we replace
1355 * the out cfg edge with a real jump.
1357 static void fix_start_block(ir_graph *irg)
1359 ir_node *initial_X = get_irg_initial_exec(irg);
1360 ir_node *start_block = get_irg_start_block(irg);
1361 ir_node *jmp = new_r_Jmp(start_block);
1363 assert(is_Proj(initial_X));
1364 exchange(initial_X, jmp);
1365 set_irg_initial_exec(irg, new_r_Bad(irg, mode_X));
1367 /* merge start block with successor if possible */
1369 foreach_out_edge(jmp, edge) {
1370 ir_node *succ = get_edge_src_irn(edge);
1371 if (!is_Block(succ))
1374 if (get_irn_arity(succ) == 1) {
1375 exchange(succ, start_block);
1383 * Modify the irg itself and the frame type.
1385 static void modify_irg(ir_graph *irg)
1387 be_abi_irg_t *env = be_get_irg_abi(irg);
1388 be_abi_call_t *call = env->call;
1389 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1390 const arch_register_t *sp = arch_env->sp;
1391 ir_type *method_type = get_entity_type(get_irg_entity(irg));
1392 be_irg_t *birg = be_birg_from_irg(irg);
1393 struct obstack *obst = be_get_be_obst(irg);
1394 be_stack_layout_t *stack_layout = be_get_irg_stack_layout(irg);
1397 ir_node *new_mem_proj;
1403 unsigned frame_size;
1406 const arch_register_t *fp_reg;
1407 ir_node *frame_pointer;
1411 ir_type *arg_type, *bet_type;
1412 lower_frame_sels_env_t ctx;
1414 DBG((dbg, LEVEL_1, "introducing abi on %+F\n", irg));
1416 old_mem = get_irg_initial_mem(irg);
1418 irp_reserve_resources(irp, IRP_RESOURCE_ENTITY_LINK);
1420 arg_type = compute_arg_type(irg, call, method_type);
1422 /* Convert the Sel nodes in the irg to frame addr nodes: */
1423 ctx.frame = get_irg_frame(irg);
1424 ctx.sp_class = arch_env->sp->reg_class;
1425 ctx.link_class = arch_env->link_class;
1426 ctx.frame_tp = get_irg_frame_type(irg);
1428 /* layout the stackframe now */
1429 if (get_type_state(ctx.frame_tp) == layout_undefined) {
1430 default_layout_compound_type(ctx.frame_tp);
1433 /* align stackframe to 4 byte */
1434 frame_size = get_type_size_bytes(ctx.frame_tp);
1435 if (frame_size % 4 != 0) {
1436 set_type_size_bytes(ctx.frame_tp, frame_size + 4 - (frame_size % 4));
1439 env->regs = pmap_create();
1441 n_params = get_method_n_params(method_type);
1442 args = OALLOCNZ(obst, ir_node*, n_params);
1444 be_add_parameter_entity_stores(irg);
1446 irg_walk_graph(irg, lower_frame_sels_walker, NULL, &ctx);
1448 irp_free_resources(irp, IRP_RESOURCE_ENTITY_LINK);
1450 /* Fill the argument vector */
1451 arg_tuple = get_irg_args(irg);
1452 foreach_out_edge(arg_tuple, edge) {
1453 ir_node *irn = get_edge_src_irn(edge);
1454 if (! is_Anchor(irn)) {
1455 int nr = get_Proj_proj(irn);
1457 DBG((dbg, LEVEL_2, "\treading arg: %d -> %+F\n", nr, irn));
1461 stack_layout->sp_relative = call->flags.bits.try_omit_fp;
1462 bet_type = call->cb->get_between_type(irg);
1463 stack_frame_init(stack_layout, arg_type, bet_type,
1464 get_irg_frame_type(irg));
1466 /* Count the register params and add them to the number of Projs for the RegParams node */
1467 for (i = 0; i < n_params; ++i) {
1468 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1469 if (arg->in_reg && args[i]) {
1470 assert(arg->reg != sp && "cannot use stack pointer as parameter register");
1471 assert(i == get_Proj_proj(args[i]));
1473 /* For now, associate the register with the old Proj from Start representing that argument. */
1474 pmap_insert(env->regs, (void *) arg->reg, args[i]);
1475 DBG((dbg, LEVEL_2, "\targ #%d -> reg %s\n", i, arg->reg->name));
1479 /* Collect all callee-save registers */
1480 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
1481 const arch_register_class_t *cls = &arch_env->register_classes[i];
1482 for (j = 0; j < cls->n_regs; ++j) {
1483 const arch_register_t *reg = &cls->regs[j];
1484 if ((reg->type & arch_register_type_state) || arch_register_is_callee_save(arch_env, reg)) {
1485 pmap_insert(env->regs, (void *) reg, NULL);
1490 fp_reg = call->flags.bits.try_omit_fp ? arch_env->sp : arch_env->bp;
1491 rbitset_clear(birg->allocatable_regs, fp_reg->global_index);
1493 /* handle start block here (place a jump in the block) */
1494 fix_start_block(irg);
1496 pmap_insert(env->regs, (void *) sp, NULL);
1497 pmap_insert(env->regs, (void *) arch_env->bp, NULL);
1498 start_bl = get_irg_start_block(irg);
1499 ir_node *const start = be_new_Start(NULL, start_bl, pmap_count(env->regs) + 1);
1500 set_irg_start(irg, start);
1503 * make proj nodes for the callee save registers.
1504 * memorize them, since Return nodes get those as inputs.
1506 * Note, that if a register corresponds to an argument, the regs map
1507 * contains the old Proj from start for that argument.
1509 rm = ALLOCAN(reg_node_map_t, pmap_count(env->regs));
1510 reg_map_to_arr(rm, env->regs);
1511 for (i = 0, n = pmap_count(env->regs); i < n; ++i) {
1512 const arch_register_t *reg = rm[i].reg;
1513 ir_mode *mode = reg->reg_class->mode;
1515 arch_register_req_type_t add_type = arch_register_req_type_none;
1519 add_type |= arch_register_req_type_produces_sp;
1520 if (!rbitset_is_set(birg->allocatable_regs, reg->global_index)) {
1521 add_type |= arch_register_req_type_ignore;
1525 proj = new_r_Proj(start, mode, nr + 1);
1526 pmap_insert(env->regs, (void *) reg, proj);
1527 be_set_constr_single_reg_out(start, nr + 1, reg, add_type);
1528 arch_set_irn_register(proj, reg);
1530 DBG((dbg, LEVEL_2, "\tregister save proj #%d -> reg %s\n", nr, reg->name));
1533 /* create a new initial memory proj */
1534 assert(is_Proj(old_mem));
1535 arch_set_irn_register_req_out(start, 0, arch_no_register_req);
1536 new_mem_proj = new_r_Proj(start, mode_M, 0);
1538 set_irg_initial_mem(irg, mem);
1540 env->init_sp = be_abi_reg_map_get(env->regs, sp);
1542 /* set new frame_pointer */
1543 frame_pointer = be_abi_reg_map_get(env->regs, fp_reg);
1544 set_irg_frame(irg, frame_pointer);
1546 /* rewire old mem users to new mem */
1547 exchange(old_mem, mem);
1549 /* keep the mem (for functions with an endless loop = no return) */
1552 set_irg_initial_mem(irg, mem);
1554 /* Now, introduce stack param nodes for all parameters passed on the stack */
1555 for (i = 0; i < n_params; ++i) {
1556 ir_node *arg_proj = args[i];
1557 ir_node *repl = NULL;
1559 if (arg_proj != NULL) {
1560 be_abi_call_arg_t *arg;
1561 ir_type *param_type;
1562 int nr = get_Proj_proj(arg_proj);
1565 nr = MIN(nr, n_params);
1566 arg = get_call_arg(call, 0, nr, 1);
1567 param_type = get_method_param_type(method_type, nr);
1570 repl = pmap_get(ir_node, env->regs, arg->reg);
1571 } else if (arg->on_stack) {
1572 ir_node *addr = be_new_FrameAddr(sp->reg_class, start_bl, frame_pointer, arg->stack_ent);
1574 /* For atomic parameters which are actually used, we create a Load node. */
1575 if (is_atomic_type(param_type) && get_irn_n_edges(args[i]) > 0) {
1576 ir_mode *mode = get_type_mode(param_type);
1577 ir_mode *load_mode = arg->load_mode;
1578 ir_node *nomem = get_irg_no_mem(irg);
1580 ir_node *load = new_r_Load(start_bl, nomem, addr, load_mode, cons_floats);
1581 repl = new_r_Proj(load, load_mode, pn_Load_res);
1583 if (mode != load_mode) {
1584 repl = new_r_Conv(start_bl, repl, mode);
1587 /* The stack parameter is not primitive (it is a struct or array),
1588 * we thus will create a node representing the parameter's address
1594 assert(repl != NULL);
1596 /* Beware: the mode of the register parameters is always the mode of the register class
1597 which may be wrong. Add Conv's then. */
1598 mode = get_irn_mode(args[i]);
1599 if (mode != get_irn_mode(repl)) {
1600 repl = new_r_Conv(get_nodes_block(repl), repl, mode);
1602 exchange(args[i], repl);
1606 /* the arg proj is not needed anymore now and should be only used by the anchor */
1607 assert(get_irn_n_edges(arg_tuple) == 1);
1608 kill_node(arg_tuple);
1609 set_irg_args(irg, new_r_Bad(irg, mode_T));
1611 /* All Return nodes hang on the End node, so look for them there. */
1612 end = get_irg_end_block(irg);
1613 for (i = 0, n = get_Block_n_cfgpreds(end); i < n; ++i) {
1614 ir_node *irn = get_Block_cfgpred(end, i);
1616 if (is_Return(irn)) {
1617 ir_node *blk = get_nodes_block(irn);
1618 ir_node *mem = get_Return_mem(irn);
1619 ir_node *ret = create_be_return(env, irn, blk, mem, get_Return_n_ress(irn));
1624 /* if we have endless loops here, n might be <= 0. Do NOT create a be_Return then,
1625 the code is dead and will never be executed. */
1628 /** Fix the state inputs of calls that still hang on unknowns */
1629 static void fix_call_state_inputs(ir_graph *irg)
1631 be_abi_irg_t *env = be_get_irg_abi(irg);
1632 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1634 arch_register_t **stateregs = NEW_ARR_F(arch_register_t*, 0);
1636 /* Collect caller save registers */
1637 n = arch_env->n_register_classes;
1638 for (i = 0; i < n; ++i) {
1640 const arch_register_class_t *cls = &arch_env->register_classes[i];
1641 for (j = 0; j < cls->n_regs; ++j) {
1642 const arch_register_t *reg = arch_register_for_index(cls, j);
1643 if (reg->type & arch_register_type_state) {
1644 ARR_APP1(arch_register_t*, stateregs, (arch_register_t *)reg);
1649 n = ARR_LEN(env->calls);
1650 n_states = ARR_LEN(stateregs);
1651 for (i = 0; i < n; ++i) {
1653 ir_node *call = env->calls[i];
1655 arity = get_irn_arity(call);
1657 /* the state reg inputs are the last n inputs of the calls */
1658 for (s = 0; s < n_states; ++s) {
1659 int inp = arity - n_states + s;
1660 const arch_register_t *reg = stateregs[s];
1661 ir_node *regnode = be_abi_reg_map_get(env->regs, reg);
1663 set_irn_n(call, inp, regnode);
1667 DEL_ARR_F(stateregs);
1671 * Create a trampoline entity for the given method.
1673 static ir_entity *create_trampoline(be_main_env_t *be, ir_entity *method)
1675 ir_type *type = get_entity_type(method);
1676 ident *old_id = get_entity_ld_ident(method);
1677 ident *id = id_mangle3("", old_id, "$stub");
1678 ir_type *parent = be->pic_trampolines_type;
1679 ir_entity *ent = new_entity(parent, old_id, type);
1680 set_entity_ld_ident(ent, id);
1681 set_entity_visibility(ent, ir_visibility_private);
1687 * Returns the trampoline entity for the given method.
1689 static ir_entity *get_trampoline(be_main_env_t *env, ir_entity *method)
1691 ir_entity *result = pmap_get(ir_entity, env->ent_trampoline_map, method);
1692 if (result == NULL) {
1693 result = create_trampoline(env, method);
1694 pmap_insert(env->ent_trampoline_map, method, result);
1700 static ir_entity *create_pic_symbol(be_main_env_t *be, ir_entity *entity)
1702 ident *old_id = get_entity_ld_ident(entity);
1703 ident *id = id_mangle3("", old_id, "$non_lazy_ptr");
1704 ir_type *e_type = get_entity_type(entity);
1705 ir_type *type = new_type_pointer(e_type);
1706 ir_type *parent = be->pic_symbols_type;
1707 ir_entity *ent = new_entity(parent, old_id, type);
1708 set_entity_ld_ident(ent, id);
1709 set_entity_visibility(ent, ir_visibility_private);
1714 static ir_entity *get_pic_symbol(be_main_env_t *env, ir_entity *entity)
1716 ir_entity *result = pmap_get(ir_entity, env->ent_pic_symbol_map, entity);
1717 if (result == NULL) {
1718 result = create_pic_symbol(env, entity);
1719 pmap_insert(env->ent_pic_symbol_map, entity, result);
1728 * Returns non-zero if a given entity can be accessed using a relative address.
1730 static int can_address_relative(ir_entity *entity)
1732 return entity_has_definition(entity) && !(get_entity_linkage(entity) & IR_LINKAGE_MERGE);
1735 static ir_node *get_pic_base(ir_graph *irg)
1737 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1738 if (arch_env->impl->get_pic_base == NULL)
1740 return arch_env->impl->get_pic_base(irg);
1743 /** patches SymConsts to work in position independent code */
1744 static void fix_pic_symconsts(ir_node *node, void *data)
1746 ir_graph *irg = get_irn_irg(node);
1747 be_main_env_t *be = be_get_irg_main_env(irg);
1757 arity = get_irn_arity(node);
1758 for (i = 0; i < arity; ++i) {
1760 ir_node *pred = get_irn_n(node, i);
1762 ir_entity *pic_symbol;
1763 ir_node *pic_symconst;
1765 if (!is_SymConst(pred))
1768 entity = get_SymConst_entity(pred);
1769 block = get_nodes_block(pred);
1771 /* calls can jump to relative addresses, so we can directly jump to
1772 the (relatively) known call address or the trampoline */
1773 if (i == 1 && is_Call(node)) {
1774 ir_entity *trampoline;
1775 ir_node *trampoline_const;
1777 if (can_address_relative(entity))
1780 dbgi = get_irn_dbg_info(pred);
1781 trampoline = get_trampoline(be, entity);
1782 trampoline_const = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
1784 set_irn_n(node, i, trampoline_const);
1788 /* everything else is accessed relative to EIP */
1789 mode = get_irn_mode(pred);
1790 pic_base = get_pic_base(irg);
1792 /* all ok now for locally constructed stuff */
1793 if (can_address_relative(entity)) {
1794 ir_node *add = new_r_Add(block, pic_base, pred, mode);
1796 /* make sure the walker doesn't visit this add again */
1797 mark_irn_visited(add);
1798 set_irn_n(node, i, add);
1802 /* get entry from pic symbol segment */
1803 dbgi = get_irn_dbg_info(pred);
1804 pic_symbol = get_pic_symbol(be, entity);
1805 pic_symconst = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
1807 add = new_r_Add(block, pic_base, pic_symconst, mode);
1808 mark_irn_visited(add);
1810 /* we need an extra indirection for global data outside our current
1811 module. The loads are always safe and can therefore float
1812 and need no memory input */
1813 load = new_r_Load(block, get_irg_no_mem(irg), add, mode, cons_floats);
1814 load_res = new_r_Proj(load, mode, pn_Load_res);
1816 set_irn_n(node, i, load_res);
1820 void be_abi_introduce(ir_graph *irg)
1822 be_abi_irg_t *env = XMALLOCZ(be_abi_irg_t);
1823 ir_node *old_frame = get_irg_frame(irg);
1824 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1825 ir_entity *entity = get_irg_entity(irg);
1826 ir_type *method_type = get_entity_type(entity);
1827 be_irg_t *birg = be_birg_from_irg(irg);
1828 struct obstack *obst = &birg->obst;
1829 ir_node *dummy = new_r_Dummy(irg,
1830 arch_env->sp->reg_class->mode);
1833 /* determine allocatable registers */
1834 assert(birg->allocatable_regs == NULL);
1835 birg->allocatable_regs = rbitset_obstack_alloc(obst, arch_env->n_registers);
1836 for (r = 0; r < arch_env->n_registers; ++r) {
1837 const arch_register_t *reg = &arch_env->registers[r];
1838 if ( !(reg->type & arch_register_type_ignore)) {
1839 rbitset_set(birg->allocatable_regs, r);
1843 /* break here if backend provides a custom API.
1844 * Note: we shouldn't have to setup any be_abi_irg_t* stuff at all,
1845 * but need more cleanup to make this work
1847 be_set_irg_abi(irg, env);
1849 be_omit_fp = be_options.omit_fp;
1851 env->keep_map = pmap_create();
1852 env->call = be_abi_call_new(arch_env->sp->reg_class);
1853 arch_env_get_call_abi(arch_env, method_type, env->call);
1855 env->init_sp = dummy;
1856 env->calls = NEW_ARR_F(ir_node*, 0);
1860 if (be_options.pic) {
1861 irg_walk_graph(irg, fix_pic_symconsts, NULL, NULL);
1864 /* Lower all call nodes in the IRG. */
1867 /* Process the IRG */
1870 /* fix call inputs for state registers */
1871 fix_call_state_inputs(irg);
1873 /* We don't need the keep map anymore. */
1874 pmap_destroy(env->keep_map);
1875 env->keep_map = NULL;
1877 /* calls array is not needed anymore */
1878 DEL_ARR_F(env->calls);
1881 /* reroute the stack origin of the calls to the true stack origin. */
1882 exchange(dummy, env->init_sp);
1883 exchange(old_frame, get_irg_frame(irg));
1885 pmap_destroy(env->regs);
1889 void be_abi_free(ir_graph *irg)
1891 be_abi_irg_t *env = be_get_irg_abi(irg);
1893 if (env->call != NULL)
1894 be_abi_call_free(env->call);
1895 assert(env->regs == NULL);
1898 be_set_irg_abi(irg, NULL);
1901 void be_put_allocatable_regs(const ir_graph *irg,
1902 const arch_register_class_t *cls, bitset_t *bs)
1904 be_irg_t *birg = be_birg_from_irg(irg);
1905 unsigned *allocatable_regs = birg->allocatable_regs;
1908 assert(bitset_size(bs) == cls->n_regs);
1909 bitset_clear_all(bs);
1910 for (i = 0; i < cls->n_regs; ++i) {
1911 const arch_register_t *reg = &cls->regs[i];
1912 if (rbitset_is_set(allocatable_regs, reg->global_index))
1917 unsigned be_get_n_allocatable_regs(const ir_graph *irg,
1918 const arch_register_class_t *cls)
1920 bitset_t *bs = bitset_alloca(cls->n_regs);
1921 be_put_allocatable_regs(irg, cls, bs);
1922 return bitset_popcount(bs);
1925 void be_set_allocatable_regs(const ir_graph *irg,
1926 const arch_register_class_t *cls,
1927 unsigned *raw_bitset)
1929 be_irg_t *birg = be_birg_from_irg(irg);
1930 unsigned *allocatable_regs = birg->allocatable_regs;
1933 rbitset_clear_all(raw_bitset, cls->n_regs);
1934 for (i = 0; i < cls->n_regs; ++i) {
1935 const arch_register_t *reg = &cls->regs[i];
1936 if (rbitset_is_set(allocatable_regs, reg->global_index))
1937 rbitset_set(raw_bitset, i);
1941 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_abi)
1942 void be_init_abi(void)
1944 FIRM_DBG_REGISTER(dbg, "firm.be.abi");