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 1: in registers, 0: on stack. */
64 unsigned callee : 1; /**< 1: someone called us. 0: We call another function */
67 const arch_register_t *reg;
70 unsigned alignment; /**< stack alignment */
71 unsigned space_before; /**< allocate space before */
72 unsigned space_after; /**< allocate space after */
75 struct be_abi_call_t {
76 be_abi_call_flags_t flags; /**< Flags describing the ABI behavior on calls */
77 int pop; /**< number of bytes the stack frame is shrinked by the callee on return. */
78 const be_abi_callbacks_t *cb;
79 ir_type *between_type;
81 const arch_register_class_t *cls_addr; /**< register class of the call address */
85 * The ABI information for the current graph.
88 be_abi_call_t *call; /**< The ABI call information. */
90 ir_node *init_sp; /**< The node representing the stack pointer
91 at the start of the function. */
93 pmap *regs; /**< A map of all callee-save and ignore regs to
94 their Projs to the RegParams node. */
95 pmap *keep_map; /**< mapping blocks to keep nodes. */
97 ir_node **calls; /**< flexible array containing all be_Call nodes */
100 static ir_heights_t *ir_heights;
102 /** Flag: if set, try to omit the frame pointer in all routines. */
103 static int be_omit_fp = 1;
105 static ir_node *be_abi_reg_map_get(pmap *map, const arch_register_t *reg)
107 return pmap_get(ir_node, map, reg);
110 static void be_abi_reg_map_set(pmap *map, const arch_register_t* reg,
113 pmap_insert(map, reg, node);
117 * Check if the given register is callee save, ie. will be saved by the callee.
119 static bool arch_register_is_callee_save(
120 const arch_env_t *arch_env,
121 const arch_register_t *reg)
123 if (arch_env->impl->register_saved_by)
124 return arch_env->impl->register_saved_by(reg, /*callee=*/1);
129 * Check if the given register is caller save, ie. must be saved by the caller.
131 static bool arch_register_is_caller_save(
132 const arch_env_t *arch_env,
133 const arch_register_t *reg)
135 if (arch_env->impl->register_saved_by)
136 return arch_env->impl->register_saved_by(reg, /*callee=*/0);
143 _ ____ ___ ____ _ _ _ _
144 / \ | __ )_ _| / ___|__ _| | | |__ __ _ ___| | _____
145 / _ \ | _ \| | | | / _` | | | '_ \ / _` |/ __| |/ / __|
146 / ___ \| |_) | | | |__| (_| | | | |_) | (_| | (__| <\__ \
147 /_/ \_\____/___| \____\__,_|_|_|_.__/ \__,_|\___|_|\_\___/
149 These callbacks are used by the backend to set the parameters
150 for a specific call type.
154 * Set compare function: compares two ABI call object arguments.
156 static int cmp_call_arg(const void *a, const void *b, size_t n)
158 const be_abi_call_arg_t *p = (const be_abi_call_arg_t*)a;
159 const be_abi_call_arg_t *q = (const be_abi_call_arg_t*)b;
161 return !(p->is_res == q->is_res && p->pos == q->pos && p->callee == q->callee);
165 * Get an ABI call object argument.
167 * @param call the abi call
168 * @param is_res true for call results, false for call arguments
169 * @param pos position of the argument
170 * @param callee context type - if we are callee or caller
172 static be_abi_call_arg_t *get_call_arg(be_abi_call_t *call, int is_res, int pos, int callee)
174 be_abi_call_arg_t arg;
177 memset(&arg, 0, sizeof(arg));
182 hash = is_res * 128 + pos;
184 return set_find(be_abi_call_arg_t, call->params, &arg, sizeof(arg), hash);
188 * Set an ABI call object argument.
190 static void remember_call_arg(be_abi_call_arg_t *arg, be_abi_call_t *call, be_abi_context_t context)
192 unsigned hash = arg->is_res * 128 + arg->pos;
193 if (context & ABI_CONTEXT_CALLEE) {
195 (void)set_insert(be_abi_call_arg_t, call->params, arg, sizeof(*arg), hash);
197 if (context & ABI_CONTEXT_CALLER) {
199 (void)set_insert(be_abi_call_arg_t, call->params, arg, sizeof(*arg), hash);
203 /* Set the flags for a call. */
204 void be_abi_call_set_flags(be_abi_call_t *call, be_abi_call_flags_t flags, const be_abi_callbacks_t *cb)
210 /* Sets the number of bytes the stackframe is shrinked by the callee on return */
211 void be_abi_call_set_pop(be_abi_call_t *call, int pop)
217 /* Set register class for call address */
218 void be_abi_call_set_call_address_reg_class(be_abi_call_t *call, const arch_register_class_t *cls)
220 call->cls_addr = cls;
224 void be_abi_call_param_stack(be_abi_call_t *call, int arg_pos,
225 ir_mode *load_mode, unsigned alignment,
226 unsigned space_before, unsigned space_after,
227 be_abi_context_t context)
229 be_abi_call_arg_t arg;
230 memset(&arg, 0, sizeof(arg));
231 assert(alignment > 0 && "Alignment must be greater than 0");
232 arg.load_mode = load_mode;
233 arg.alignment = alignment;
234 arg.space_before = space_before;
235 arg.space_after = space_after;
239 remember_call_arg(&arg, call, context);
242 void be_abi_call_param_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
244 be_abi_call_arg_t arg;
245 memset(&arg, 0, sizeof(arg));
252 remember_call_arg(&arg, call, context);
255 void be_abi_call_res_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
257 be_abi_call_arg_t arg;
258 memset(&arg, 0, sizeof(arg));
265 remember_call_arg(&arg, call, context);
268 /* Get the flags of a ABI call object. */
269 be_abi_call_flags_t be_abi_call_get_flags(const be_abi_call_t *call)
275 * Constructor for a new ABI call object.
277 * @param cls_addr register class of the call address
279 * @return the new ABI call object
281 static be_abi_call_t *be_abi_call_new(const arch_register_class_t *cls_addr)
283 be_abi_call_t *call = XMALLOCZ(be_abi_call_t);
286 call->params = new_set(cmp_call_arg, 16);
288 call->cls_addr = cls_addr;
290 call->flags.bits.try_omit_fp = be_omit_fp;
296 * Destructor for an ABI call object.
298 static void be_abi_call_free(be_abi_call_t *call)
300 del_set(call->params);
305 * Initializes the frame layout from parts
307 * @param frame the stack layout that will be initialized
308 * @param args the stack argument layout type
309 * @param between the between layout type
310 * @param locals the method frame type
312 * @return the initialized stack layout
314 static be_stack_layout_t *stack_frame_init(be_stack_layout_t *frame, ir_type *args,
315 ir_type *between, ir_type *locals)
317 frame->arg_type = args;
318 frame->between_type = between;
319 frame->frame_type = locals;
320 frame->initial_offset = 0;
321 frame->initial_bias = 0;
322 frame->order[1] = between;
324 /* typical decreasing stack: locals have the
325 * lowest addresses, arguments the highest */
326 frame->order[0] = locals;
327 frame->order[2] = args;
338 Adjustment of the calls inside a graph.
343 * Transform a call node into a be_Call node.
345 * @param env The ABI environment for the current irg.
346 * @param irn The call node.
347 * @param curr_sp The stack pointer node to use.
348 * @return The stack pointer after the call.
350 static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
352 ir_graph *irg = get_irn_irg(irn);
353 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
354 ir_type *call_tp = get_Call_type(irn);
355 ir_node *call_ptr = get_Call_ptr(irn);
356 size_t n_params = get_method_n_params(call_tp);
357 ir_node *curr_mem = get_Call_mem(irn);
358 ir_node *bl = get_nodes_block(irn);
360 const arch_register_t *sp = arch_env->sp;
361 be_abi_call_t *call = be_abi_call_new(sp->reg_class);
362 ir_mode *mach_mode = sp->reg_class->mode;
363 int n_res = get_method_n_ress(call_tp);
365 ir_node *res_proj = NULL;
366 int n_reg_params = 0;
367 int n_stack_params = 0;
370 const arch_register_t **states = NEW_ARR_F(const arch_register_t*, 0);
371 const arch_register_t **destroyed_regs = NEW_ARR_F(const arch_register_t*, 0);
375 int n_reg_results = 0;
377 int *stack_param_idx;
379 int throws_exception;
384 /* Let the isa fill out the abi description for that call node. */
385 arch_env_get_call_abi(arch_env, call_tp, call);
387 /* Insert code to put the stack arguments on the stack. */
388 assert((size_t)get_Call_n_params(irn) == n_params);
389 stack_param_idx = ALLOCAN(int, n_params);
390 for (p = 0; p < n_params; ++p) {
391 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
394 int arg_size = get_type_size_bytes(get_method_param_type(call_tp, p));
396 stack_size += round_up2(arg->space_before, arg->alignment);
397 stack_size += round_up2(arg_size, arg->alignment);
398 stack_size += round_up2(arg->space_after, arg->alignment);
400 stack_param_idx[n_stack_params++] = p;
404 /* Collect all arguments which are passed in registers. */
405 reg_param_idxs = ALLOCAN(int, n_params);
406 for (p = 0; p < n_params; ++p) {
407 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
408 if (arg && arg->in_reg) {
409 reg_param_idxs[n_reg_params++] = p;
414 * If the stack is decreasing and we do not want to store sequentially,
415 * or someone else allocated the call frame
416 * we allocate as much space on the stack all parameters need, by
417 * moving the stack pointer along the stack's direction.
419 * Note: we also have to do this for stack_size == 0, because we may have
420 * to adjust stack alignment for the call.
422 curr_sp = be_new_IncSP(sp, bl, curr_sp, stack_size, 1);
424 dbgi = get_irn_dbg_info(irn);
425 /* If there are some parameters which shall be passed on the stack. */
426 if (n_stack_params > 0) {
428 ir_node **in = ALLOCAN(ir_node*, n_stack_params+1);
431 curr_mem = get_Call_mem(irn);
432 in[n_in++] = curr_mem;
434 for (i = 0; i < n_stack_params; ++i) {
435 int p = stack_param_idx[i];
436 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
437 ir_node *param = get_Call_param(irn, p);
438 ir_node *addr = curr_sp;
440 ir_type *param_type = get_method_param_type(call_tp, p);
441 int param_size = get_type_size_bytes(param_type) + arg->space_after;
444 * If we wanted to build the arguments sequentially,
445 * the stack pointer for the next must be incremented,
446 * and the memory value propagated.
448 curr_ofs += arg->space_before;
449 curr_ofs = round_up2(curr_ofs, arg->alignment);
451 /* Make the expression to compute the argument's offset. */
453 ir_mode *constmode = mach_mode;
454 if (mode_is_reference(mach_mode)) {
457 addr = new_r_Const_long(irg, constmode, curr_ofs);
458 addr = new_r_Add(bl, curr_sp, addr, mach_mode);
461 /* Insert a store for primitive arguments. */
462 if (is_atomic_type(param_type)) {
463 ir_node *nomem = get_irg_no_mem(irg);
464 ir_node *mem_input = nomem;
465 ir_node *store = new_rd_Store(dbgi, bl, mem_input, addr, param, cons_none);
466 mem = new_r_Proj(store, mode_M, pn_Store_M);
468 /* Make a mem copy for compound arguments. */
471 assert(mode_is_reference(get_irn_mode(param)));
472 copy = new_rd_CopyB(dbgi, bl, curr_mem, addr, param, param_type);
473 mem = new_r_Proj(copy, mode_M, pn_CopyB_M);
476 curr_ofs += param_size;
481 /* We need the sync only, if we didn't build the stores sequentially. */
482 if (n_stack_params >= 1) {
483 curr_mem = new_r_Sync(bl, n_in, in);
485 curr_mem = get_Call_mem(irn);
489 /* Put caller save into the destroyed set and state registers in the states
491 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
493 const arch_register_class_t *cls = &arch_env->register_classes[i];
494 for (j = 0; j < cls->n_regs; ++j) {
495 const arch_register_t *reg = arch_register_for_index(cls, j);
497 /* even if destroyed all is specified, neither SP nor FP are
498 * destroyed (else bad things will happen) */
499 if (reg == arch_env->sp || reg == arch_env->bp)
502 if (reg->type & arch_register_type_state) {
503 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
504 ARR_APP1(const arch_register_t*, states, reg);
505 /* we're already in the destroyed set so no need for further
509 if (arch_register_is_caller_save(arch_env, reg)) {
510 if (!(reg->type & arch_register_type_ignore)) {
511 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
517 /* search the largest result proj number */
518 res_projs = ALLOCANZ(ir_node*, n_res);
520 foreach_out_edge(irn, edge) {
521 ir_node *irn = get_edge_src_irn(edge);
523 if (!is_Proj(irn) || get_Proj_proj(irn) != pn_Call_T_result)
526 foreach_out_edge(irn, res_edge) {
528 ir_node *res = get_edge_src_irn(res_edge);
530 assert(is_Proj(res));
532 proj = get_Proj_proj(res);
533 assert(proj < n_res);
534 assert(res_projs[proj] == NULL);
535 res_projs[proj] = res;
541 /** TODO: this is not correct for cases where return values are passed
542 * on the stack, but no known ABI does this currently...
544 n_reg_results = n_res;
547 in = ALLOCAN(ir_node*, n_reg_params + ARR_LEN(states));
549 /* make the back end call node and set its register requirements. */
550 for (i = 0; i < n_reg_params; ++i) {
551 in[n_ins++] = get_Call_param(irn, reg_param_idxs[i]);
554 /* add state registers ins */
555 for (s = 0; s < ARR_LEN(states); ++s) {
556 const arch_register_t *reg = states[s];
557 const arch_register_class_t *cls = reg->reg_class;
558 ir_node *regnode = new_r_Unknown(irg, cls->mode);
559 in[n_ins++] = regnode;
561 assert(n_ins == (int) (n_reg_params + ARR_LEN(states)));
563 /* ins collected, build the call */
564 throws_exception = ir_throws_exception(irn);
565 if (env->call->flags.bits.call_has_imm && is_SymConst(call_ptr)) {
567 low_call = be_new_Call(dbgi, irg, bl, curr_mem, sp->single_req, curr_sp,
568 sp->single_req, curr_sp,
569 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
570 n_ins, in, get_Call_type(irn));
571 be_Call_set_entity(low_call, get_SymConst_entity(call_ptr));
574 low_call = be_new_Call(dbgi, irg, bl, curr_mem, sp->single_req, curr_sp,
575 call->cls_addr->class_req, call_ptr,
576 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
577 n_ins, in, get_Call_type(irn));
579 ir_set_throws_exception(low_call, throws_exception);
580 be_Call_set_pop(low_call, call->pop);
582 /* put the call into the list of all calls for later processing */
583 ARR_APP1(ir_node *, env->calls, low_call);
585 /* create new stack pointer */
586 curr_sp = new_r_Proj(low_call, get_irn_mode(curr_sp), pn_be_Call_sp);
587 be_set_constr_single_reg_out(low_call, pn_be_Call_sp, sp,
588 arch_register_req_type_ignore | arch_register_req_type_produces_sp);
589 arch_set_irn_register(curr_sp, sp);
591 /* now handle results */
592 for (i = 0; i < n_res; ++i) {
593 ir_node *proj = res_projs[i];
594 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
596 /* returns values on stack not supported yet */
600 shift the proj number to the right, since we will drop the
601 unspeakable Proj_T from the Call. Therefore, all real argument
602 Proj numbers must be increased by pn_be_Call_first_res
604 long pn = i + pn_be_Call_first_res;
607 ir_type *res_type = get_method_res_type(call_tp, i);
608 ir_mode *mode = get_type_mode(res_type);
609 proj = new_r_Proj(low_call, mode, pn);
612 set_Proj_pred(proj, low_call);
613 set_Proj_proj(proj, pn);
617 /* remove register from destroyed regs */
619 size_t n = ARR_LEN(destroyed_regs);
620 for (j = 0; j < n; ++j) {
621 if (destroyed_regs[j] == arg->reg) {
622 destroyed_regs[j] = destroyed_regs[n-1];
623 ARR_SHRINKLEN(destroyed_regs,n-1);
630 DBG((dbg, LEVEL_3, "\tcreated backend call %+F\n", low_call));
632 /* Set the register classes and constraints of the Call parameters. */
633 for (i = 0; i < n_reg_params; ++i) {
634 int index = reg_param_idxs[i];
635 be_abi_call_arg_t *arg = get_call_arg(call, 0, index, 0);
636 assert(arg->reg != NULL);
638 be_set_constr_single_reg_in(low_call, n_be_Call_first_arg + i,
639 arg->reg, arch_register_req_type_none);
642 /* Set the register constraints of the results. */
643 for (i = 0; i < n_res; ++i) {
644 ir_node *proj = res_projs[i];
645 const be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
646 int pn = get_Proj_proj(proj);
649 be_set_constr_single_reg_out(low_call, pn, arg->reg,
650 arch_register_req_type_none);
651 arch_set_irn_register(proj, arg->reg);
653 exchange(irn, low_call);
655 /* kill the ProjT node */
656 if (res_proj != NULL) {
660 /* Make additional projs for the caller save registers
661 and the Keep node which keeps them alive. */
667 int curr_res_proj = pn_be_Call_first_res + n_reg_results;
670 n_ins = ARR_LEN(destroyed_regs) + n_reg_results + 1;
671 in = ALLOCAN(ir_node *, n_ins);
673 /* also keep the stack pointer */
674 set_irn_link(curr_sp, (void*) sp);
677 for (d = 0; d < ARR_LEN(destroyed_regs); ++d) {
678 const arch_register_t *reg = destroyed_regs[d];
679 ir_node *proj = new_r_Proj(low_call, reg->reg_class->mode, curr_res_proj);
681 /* memorize the register in the link field. we need afterwards to set the register class of the keep correctly. */
682 be_set_constr_single_reg_out(low_call, curr_res_proj, reg,
683 arch_register_req_type_none);
684 arch_set_irn_register(proj, reg);
686 set_irn_link(proj, (void*) reg);
691 for (i = 0; i < n_reg_results; ++i) {
692 ir_node *proj = res_projs[i];
693 const arch_register_t *reg = arch_get_irn_register(proj);
694 set_irn_link(proj, (void*) reg);
699 /* create the Keep for the caller save registers */
700 keep = be_new_Keep(bl, n, in);
701 for (i = 0; i < n; ++i) {
702 const arch_register_t *reg = (const arch_register_t*)get_irn_link(in[i]);
703 be_node_set_reg_class_in(keep, i, reg->reg_class);
707 /* Clean up the stack. */
708 assert(stack_size >= call->pop);
709 stack_size -= call->pop;
711 if (stack_size > 0) {
712 ir_node *mem_proj = NULL;
714 foreach_out_edge(low_call, edge) {
715 ir_node *irn = get_edge_src_irn(edge);
716 if (is_Proj(irn) && get_Proj_proj(irn) == pn_Call_M) {
723 mem_proj = new_r_Proj(low_call, mode_M, pn_be_Call_M);
724 keep_alive(mem_proj);
727 /* Clean up the stack frame or revert alignment fixes if we allocated it */
728 curr_sp = be_new_IncSP(sp, bl, curr_sp, -stack_size, 0);
730 be_abi_call_free(call);
733 DEL_ARR_F(destroyed_regs);
739 * Adjust the size of a node representing a stack alloc or free for the minimum stack alignment.
741 * @param alignment the minimum stack alignment
742 * @param size the node containing the non-aligned size
743 * @param block the block where new nodes are allocated on
744 * @param dbg debug info for new nodes
746 * @return a node representing the aligned size
748 static ir_node *adjust_alloc_size(unsigned stack_alignment, ir_node *size,
749 ir_node *block, dbg_info *dbg)
751 if (stack_alignment > 1) {
757 assert(is_po2(stack_alignment));
759 mode = get_irn_mode(size);
760 tv = new_tarval_from_long(stack_alignment-1, mode);
761 irg = get_Block_irg(block);
762 mask = new_r_Const(irg, tv);
763 size = new_rd_Add(dbg, block, size, mask, mode);
765 tv = new_tarval_from_long(-(long)stack_alignment, mode);
766 mask = new_r_Const(irg, tv);
767 size = new_rd_And(dbg, block, size, mask, mode);
773 * The alloca is transformed into a back end alloca node and connected to the stack nodes.
775 static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp)
777 ir_node *block = get_nodes_block(alloc);
778 ir_graph *irg = get_Block_irg(block);
779 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
780 ir_node *alloc_mem = NULL;
781 ir_node *alloc_res = NULL;
782 ir_type *type = get_Alloc_type(alloc);
789 unsigned stack_alignment;
791 /* all non-stack Alloc nodes should already be lowered before the backend */
792 assert(get_Alloc_where(alloc) == stack_alloc);
794 foreach_out_edge(alloc, edge) {
795 ir_node *irn = get_edge_src_irn(edge);
797 assert(is_Proj(irn));
798 switch (get_Proj_proj(irn)) {
810 /* Beware: currently Alloc nodes without a result might happen,
811 only escape analysis kills them and this phase runs only for object
812 oriented source. We kill the Alloc here. */
813 if (alloc_res == NULL && alloc_mem) {
814 exchange(alloc_mem, get_Alloc_mem(alloc));
818 dbg = get_irn_dbg_info(alloc);
819 count = get_Alloc_count(alloc);
821 /* we might need to multiply the count with the element size */
822 if (!is_unknown_type(type) && get_type_size_bytes(type) != 1) {
823 ir_mode *mode = get_irn_mode(count);
824 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type),
826 ir_node *cnst = new_rd_Const(dbg, irg, tv);
827 size = new_rd_Mul(dbg, block, count, cnst, mode);
832 /* The stack pointer will be modified in an unknown manner.
833 We cannot omit it. */
834 env->call->flags.bits.try_omit_fp = 0;
836 stack_alignment = 1 << arch_env->stack_alignment;
837 size = adjust_alloc_size(stack_alignment, size, block, dbg);
838 new_alloc = be_new_AddSP(arch_env->sp, block, curr_sp, size);
839 set_irn_dbg_info(new_alloc, dbg);
841 if (alloc_mem != NULL) {
845 addsp_mem = new_r_Proj(new_alloc, mode_M, pn_be_AddSP_M);
847 /* We need to sync the output mem of the AddSP with the input mem
848 edge into the alloc node. */
849 ins[0] = get_Alloc_mem(alloc);
851 sync = new_r_Sync(block, 2, ins);
853 exchange(alloc_mem, sync);
856 exchange(alloc, new_alloc);
858 /* fix projnum of alloca res */
859 set_Proj_proj(alloc_res, pn_be_AddSP_res);
861 curr_sp = new_r_Proj(new_alloc, get_irn_mode(curr_sp), pn_be_AddSP_sp);
868 * The Free is transformed into a back end free node and connected to the stack nodes.
870 static ir_node *adjust_free(be_abi_irg_t *env, ir_node *free, ir_node *curr_sp)
872 ir_node *block = get_nodes_block(free);
873 ir_graph *irg = get_irn_irg(free);
874 ir_type *type = get_Free_type(free);
875 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
876 ir_mode *sp_mode = arch_env->sp->reg_class->mode;
877 dbg_info *dbg = get_irn_dbg_info(free);
878 ir_node *subsp, *mem, *res, *size, *sync;
880 unsigned stack_alignment;
882 /* all non-stack-alloc Free nodes should already be lowered before the
884 assert(get_Free_where(free) == stack_alloc);
886 /* we might need to multiply the size with the element size */
887 if (!is_unknown_type(type) && get_type_size_bytes(type) != 1) {
888 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type), mode_Iu);
889 ir_node *cnst = new_rd_Const(dbg, irg, tv);
890 ir_node *mul = new_rd_Mul(dbg, block, get_Free_count(free),
894 size = get_Free_count(free);
897 stack_alignment = 1 << arch_env->stack_alignment;
898 size = adjust_alloc_size(stack_alignment, size, block, dbg);
900 /* The stack pointer will be modified in an unknown manner.
901 We cannot omit it. */
902 env->call->flags.bits.try_omit_fp = 0;
903 subsp = be_new_SubSP(arch_env->sp, block, curr_sp, size);
904 set_irn_dbg_info(subsp, dbg);
906 mem = new_r_Proj(subsp, mode_M, pn_be_SubSP_M);
907 res = new_r_Proj(subsp, sp_mode, pn_be_SubSP_sp);
909 /* we need to sync the memory */
910 in[0] = get_Free_mem(free);
912 sync = new_r_Sync(block, 2, in);
914 /* and make the AddSP dependent on the former memory */
915 add_irn_dep(subsp, get_Free_mem(free));
918 exchange(free, sync);
925 * Check if a node is somehow data dependent on another one.
926 * both nodes must be in the same basic block.
927 * @param n1 The first node.
928 * @param n2 The second node.
929 * @return 1, if n1 is data dependent (transitively) on n2, 0 if not.
931 static int dependent_on(ir_node *n1, ir_node *n2)
933 assert(get_nodes_block(n1) == get_nodes_block(n2));
935 return heights_reachable_in_block(ir_heights, n1, n2);
938 static int cmp_call_dependency(const void *c1, const void *c2)
940 ir_node *n1 = *(ir_node **) c1;
941 ir_node *n2 = *(ir_node **) c2;
945 Classical qsort() comparison function behavior:
946 0 if both elements are equal
947 1 if second is "smaller" that first
948 -1 if first is "smaller" that second
950 if (dependent_on(n1, n2))
953 if (dependent_on(n2, n1))
956 /* The nodes have no depth order, but we need a total order because qsort()
959 * Additionally, we need to respect transitive dependencies. Consider a
960 * Call a depending on Call b and an independent Call c.
961 * We MUST NOT order c > a and b > c. */
962 h1 = get_irn_height(ir_heights, n1);
963 h2 = get_irn_height(ir_heights, n2);
964 if (h1 < h2) return -1;
965 if (h1 > h2) return 1;
966 /* Same height, so use a random (but stable) order */
967 return get_irn_idx(n1) - get_irn_idx(n2);
971 * Walker: links all Call/Alloc/Free nodes to the Block they are contained.
973 static void link_ops_in_block_walker(ir_node *irn, void *data)
975 be_abi_irg_t *env = (be_abi_irg_t*)data;
976 unsigned code = get_irn_opcode(irn);
978 if (code == iro_Call ||
979 (code == iro_Alloc && get_Alloc_where(irn) == stack_alloc) ||
980 (code == iro_Free && get_Free_where(irn) == stack_alloc)) {
981 ir_node *bl = get_nodes_block(irn);
982 void *save = get_irn_link(bl);
984 set_irn_link(irn, save);
985 set_irn_link(bl, irn);
988 if (code == iro_Builtin && get_Builtin_kind(irn) == ir_bk_return_address) {
989 ir_node *param = get_Builtin_param(irn, 0);
990 ir_tarval *tv = get_Const_tarval(param);
991 unsigned long value = get_tarval_long(tv);
992 /* use ebp, so the climbframe algo works... */
994 env->call->flags.bits.try_omit_fp = 0;
1001 * Process all Call/Alloc/Free nodes inside a basic block.
1002 * Note that the link field of the block must contain a linked list of all
1003 * nodes inside the Block. We first order this list according to data dependency
1004 * and that connect the nodes together.
1006 static void process_ops_in_block(ir_node *bl, void *data)
1008 be_abi_irg_t *env = (be_abi_irg_t*)data;
1009 ir_node *curr_sp = env->init_sp;
1016 for (irn = (ir_node*)get_irn_link(bl); irn != NULL;
1017 irn = (ir_node*)get_irn_link(irn)) {
1021 nodes = ALLOCAN(ir_node*, n_nodes);
1022 for (irn = (ir_node*)get_irn_link(bl), n = 0; irn != NULL;
1023 irn = (ir_node*)get_irn_link(irn), ++n) {
1027 /* If there were call nodes in the block. */
1032 /* order the call nodes according to data dependency */
1033 qsort(nodes, n_nodes, sizeof(nodes[0]), cmp_call_dependency);
1035 for (i = n_nodes - 1; i >= 0; --i) {
1036 ir_node *irn = nodes[i];
1038 DBG((dbg, LEVEL_3, "\tprocessing call %+F\n", irn));
1039 switch (get_irn_opcode(irn)) {
1042 /* The stack pointer will be modified due to a call. */
1043 env->call->flags.bits.try_omit_fp = 0;
1045 curr_sp = adjust_call(env, irn, curr_sp);
1048 if (get_Alloc_where(irn) == stack_alloc)
1049 curr_sp = adjust_alloc(env, irn, curr_sp);
1052 if (get_Free_where(irn) == stack_alloc)
1053 curr_sp = adjust_free(env, irn, curr_sp);
1056 panic("invalid call");
1060 /* Keep the last stack state in the block by tying it to Keep node,
1061 * the proj from calls is already kept */
1062 if (curr_sp != env->init_sp &&
1063 !(is_Proj(curr_sp) && be_is_Call(get_Proj_pred(curr_sp)))) {
1065 keep = be_new_Keep(bl, 1, nodes);
1066 pmap_insert(env->keep_map, bl, keep);
1070 set_irn_link(bl, curr_sp);
1074 * Adjust all call nodes in the graph to the ABI conventions.
1076 static void process_calls(ir_graph *const irg, be_abi_irg_t *const abi)
1078 irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, abi);
1080 ir_heights = heights_new(irg);
1081 irg_block_walk_graph(irg, NULL, process_ops_in_block, abi);
1082 heights_free(ir_heights);
1086 * Computes the stack argument layout type.
1087 * Changes a possibly allocated value param type by moving
1088 * entities to the stack layout type.
1090 * @param call the current call ABI
1091 * @param method_type the method type
1093 * @return the stack argument layout type
1095 static ir_type *compute_arg_type(ir_graph *irg, be_abi_call_t *call,
1096 ir_type *method_type)
1098 struct obstack *obst = be_get_be_obst(irg);
1099 ir_type *frame_type = get_irg_frame_type(irg);
1100 size_t n_params = get_method_n_params(method_type);
1101 size_t n_frame_members = get_compound_n_members(frame_type);
1102 ir_entity *va_start_entity = NULL;
1108 ir_entity **map = OALLOCNZ(obst, ir_entity*, n_params);
1109 res = new_type_struct(new_id_from_chars("arg_type", 8));
1111 /* collect existing entities for value_param_types */
1112 for (f = n_frame_members; f > 0; ) {
1113 ir_entity *entity = get_compound_member(frame_type, --f);
1116 set_entity_link(entity, NULL);
1117 if (!is_parameter_entity(entity))
1119 num = get_entity_parameter_number(entity);
1120 if (num == IR_VA_START_PARAMETER_NUMBER) {
1121 /* move entity to new arg_type */
1122 set_entity_owner(entity, res);
1123 va_start_entity = entity;
1126 assert(num < n_params);
1127 if (map[num] != NULL)
1128 panic("multiple entities for parameter %u in %+F found", f, irg);
1130 if (num != n_params && get_call_arg(call, 0, num, 1)->in_reg) {
1131 /* don't move this entity */
1136 /* move entity to new arg_type */
1137 set_entity_owner(entity, res);
1140 for (i = 0; i < n_params; ++i) {
1141 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1142 ir_type *param_type = get_method_param_type(method_type, i);
1148 if (entity == NULL) {
1149 /* create a new entity */
1150 entity = new_parameter_entity(res, i, param_type);
1152 ofs += arg->space_before;
1153 ofs = round_up2(ofs, arg->alignment);
1154 set_entity_offset(entity, ofs);
1155 ofs += arg->space_after;
1156 ofs += get_type_size_bytes(param_type);
1157 arg->stack_ent = entity;
1159 if (va_start_entity != NULL) {
1160 set_entity_offset(va_start_entity, ofs);
1162 set_type_size_bytes(res, ofs);
1163 set_type_state(res, layout_fixed);
1169 const arch_register_t *reg;
1173 static int cmp_regs(const void *a, const void *b)
1175 const reg_node_map_t *p = (const reg_node_map_t*)a;
1176 const reg_node_map_t *q = (const reg_node_map_t*)b;
1178 if (p->reg->reg_class == q->reg->reg_class)
1179 return p->reg->index - q->reg->index;
1181 return p->reg->reg_class < q->reg->reg_class ? -1 : +1;
1184 static void reg_map_to_arr(reg_node_map_t *res, pmap *reg_map)
1187 size_t n = pmap_count(reg_map);
1190 foreach_pmap(reg_map, ent) {
1191 res[i].reg = (const arch_register_t*)ent->key;
1192 res[i].irn = (ir_node*)ent->value;
1196 qsort(res, n, sizeof(res[0]), cmp_regs);
1200 * Creates a be_Return for a Return node.
1202 * @param @env the abi environment
1203 * @param irn the Return node or NULL if there was none
1204 * @param bl the block where the be_Retun should be placed
1205 * @param mem the current memory
1206 * @param n_res number of return results
1208 static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl,
1209 ir_node *mem, int n_res)
1211 be_abi_call_t *call = env->call;
1212 ir_graph *irg = get_Block_irg(bl);
1213 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1215 pmap *reg_map = pmap_create();
1216 ir_node *keep = pmap_get(ir_node, env->keep_map, bl);
1223 const arch_register_t **regs;
1227 get the valid stack node in this block.
1228 If we had a call in that block there is a Keep constructed by process_calls()
1229 which points to the last stack modification in that block. we'll use
1230 it then. Else we use the stack from the start block and let
1231 the ssa construction fix the usage.
1233 stack = be_abi_reg_map_get(env->regs, arch_env->sp);
1235 stack = get_irn_n(keep, 0);
1237 remove_End_keepalive(get_irg_end(irg), keep);
1240 /* Insert results for Return into the register map. */
1241 for (i = 0; i < n_res; ++i) {
1242 ir_node *res = get_Return_res(irn, i);
1243 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1244 assert(arg->in_reg && "return value must be passed in register");
1245 pmap_insert(reg_map, (void *) arg->reg, res);
1248 /* Add uses of the callee save registers. */
1249 foreach_pmap(env->regs, ent) {
1250 const arch_register_t *reg = (const arch_register_t*)ent->key;
1251 if ((reg->type & arch_register_type_ignore) || arch_register_is_callee_save(arch_env, reg))
1252 pmap_insert(reg_map, ent->key, ent->value);
1255 be_abi_reg_map_set(reg_map, arch_env->sp, stack);
1258 Maximum size of the in array for Return nodes is
1259 return args + callee save/ignore registers + memory + stack pointer
1261 in_max = pmap_count(reg_map) + n_res + 2;
1263 in = ALLOCAN(ir_node*, in_max);
1264 regs = ALLOCAN(arch_register_t const*, in_max);
1267 in[1] = be_abi_reg_map_get(reg_map, arch_env->sp);
1269 regs[1] = arch_env->sp;
1272 /* clear SP entry, since it has already been grown. */
1273 pmap_insert(reg_map, (void *) arch_env->sp, NULL);
1274 for (i = 0; i < n_res; ++i) {
1275 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1277 in[n] = be_abi_reg_map_get(reg_map, arg->reg);
1278 regs[n++] = arg->reg;
1280 /* Clear the map entry to mark the register as processed. */
1281 be_abi_reg_map_set(reg_map, arg->reg, NULL);
1284 /* grow the rest of the stuff. */
1285 foreach_pmap(reg_map, ent) {
1287 in[n] = (ir_node*)ent->value;
1288 regs[n++] = (const arch_register_t*)ent->key;
1292 /* The in array for the new back end return is now ready. */
1294 dbgi = get_irn_dbg_info(irn);
1298 /* we have to pop the shadow parameter in in case of struct returns */
1300 ret = be_new_Return(dbgi, irg, bl, n_res, pop, n, in);
1302 /* Set the register classes of the return's parameter accordingly. */
1303 for (i = 0; i < n; ++i) {
1304 if (regs[i] == NULL)
1307 be_set_constr_single_reg_in(ret, i, regs[i], arch_register_req_type_none);
1310 /* Free the space of the Epilog's in array and the register <-> proj map. */
1311 pmap_destroy(reg_map);
1316 typedef struct lower_frame_sels_env_t {
1317 ir_node *frame; /**< the current frame */
1318 const arch_register_class_t *sp_class; /**< register class of the stack pointer */
1319 } lower_frame_sels_env_t;
1322 * Walker: Replaces Sels of frame type and
1323 * value param type entities by FrameAddress.
1324 * Links all used entities.
1326 static void lower_frame_sels_walker(ir_node *irn, void *data)
1328 lower_frame_sels_env_t *ctx = (lower_frame_sels_env_t*)data;
1331 ir_node *ptr = get_Sel_ptr(irn);
1333 if (ptr == ctx->frame) {
1334 ir_entity *ent = get_Sel_entity(irn);
1335 ir_node *bl = get_nodes_block(irn);
1338 nw = be_new_FrameAddr(ctx->sp_class, bl, ctx->frame, ent);
1345 * The start block has no jump, instead it has an initial exec Proj.
1346 * The backend wants to handle all blocks the same way, so we replace
1347 * the out cfg edge with a real jump.
1349 static void fix_start_block(ir_graph *irg)
1351 ir_node *initial_X = get_irg_initial_exec(irg);
1352 ir_node *start_block = get_irg_start_block(irg);
1353 ir_node *jmp = new_r_Jmp(start_block);
1355 assert(is_Proj(initial_X));
1356 exchange(initial_X, jmp);
1357 set_irg_initial_exec(irg, new_r_Bad(irg, mode_X));
1359 /* merge start block with successor if possible */
1361 foreach_out_edge(jmp, edge) {
1362 ir_node *succ = get_edge_src_irn(edge);
1363 if (!is_Block(succ))
1366 if (get_irn_arity(succ) == 1) {
1367 exchange(succ, start_block);
1375 * Modify the irg itself and the frame type.
1377 static void modify_irg(ir_graph *const irg, be_abi_irg_t *const env)
1379 be_abi_call_t *call = env->call;
1380 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1381 const arch_register_t *sp = arch_env->sp;
1382 ir_type *method_type = get_entity_type(get_irg_entity(irg));
1383 be_irg_t *birg = be_birg_from_irg(irg);
1384 struct obstack *obst = be_get_be_obst(irg);
1385 be_stack_layout_t *stack_layout = be_get_irg_stack_layout(irg);
1388 ir_node *new_mem_proj;
1396 const arch_register_t *fp_reg;
1397 ir_node *frame_pointer;
1401 ir_type *arg_type, *bet_type;
1402 lower_frame_sels_env_t ctx;
1404 DBG((dbg, LEVEL_1, "introducing abi on %+F\n", irg));
1406 old_mem = get_irg_initial_mem(irg);
1408 irp_reserve_resources(irp, IRP_RESOURCE_ENTITY_LINK);
1410 arg_type = compute_arg_type(irg, call, method_type);
1412 /* Convert the Sel nodes in the irg to frame addr nodes: */
1413 ctx.frame = get_irg_frame(irg);
1414 ctx.sp_class = arch_env->sp->reg_class;
1416 ir_type *const frame_tp = get_irg_frame_type(irg);
1417 /* layout the stackframe now */
1418 if (get_type_state(frame_tp) == layout_undefined) {
1419 default_layout_compound_type(frame_tp);
1422 /* align stackframe */
1423 unsigned const alignment = 1U << arch_env->stack_alignment;
1424 unsigned const frame_size = round_up2(get_type_size_bytes(frame_tp), alignment);
1425 set_type_size_bytes(frame_tp, frame_size);
1427 env->regs = pmap_create();
1429 n_params = get_method_n_params(method_type);
1430 args = OALLOCNZ(obst, ir_node*, n_params);
1432 be_add_parameter_entity_stores(irg);
1434 irg_walk_graph(irg, lower_frame_sels_walker, NULL, &ctx);
1436 irp_free_resources(irp, IRP_RESOURCE_ENTITY_LINK);
1438 /* Fill the argument vector */
1439 arg_tuple = get_irg_args(irg);
1440 foreach_out_edge(arg_tuple, edge) {
1441 ir_node *irn = get_edge_src_irn(edge);
1442 if (! is_Anchor(irn)) {
1443 int nr = get_Proj_proj(irn);
1445 DBG((dbg, LEVEL_2, "\treading arg: %d -> %+F\n", nr, irn));
1449 stack_layout->sp_relative = call->flags.bits.try_omit_fp;
1450 bet_type = call->cb->get_between_type(irg);
1451 stack_frame_init(stack_layout, arg_type, bet_type,
1452 get_irg_frame_type(irg));
1454 /* Count the register params and add them to the number of Projs for the RegParams node */
1455 for (i = 0; i < n_params; ++i) {
1456 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1457 if (arg->in_reg && args[i]) {
1458 assert(arg->reg != sp && "cannot use stack pointer as parameter register");
1459 assert(i == get_Proj_proj(args[i]));
1461 /* For now, associate the register with the old Proj from Start representing that argument. */
1462 pmap_insert(env->regs, (void *) arg->reg, args[i]);
1463 DBG((dbg, LEVEL_2, "\targ #%d -> reg %s\n", i, arg->reg->name));
1467 /* Collect all callee-save registers */
1468 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
1469 const arch_register_class_t *cls = &arch_env->register_classes[i];
1470 for (j = 0; j < cls->n_regs; ++j) {
1471 const arch_register_t *reg = &cls->regs[j];
1472 if ((reg->type & arch_register_type_state) || arch_register_is_callee_save(arch_env, reg)) {
1473 pmap_insert(env->regs, (void *) reg, NULL);
1478 fp_reg = call->flags.bits.try_omit_fp ? arch_env->sp : arch_env->bp;
1479 rbitset_clear(birg->allocatable_regs, fp_reg->global_index);
1481 /* handle start block here (place a jump in the block) */
1482 fix_start_block(irg);
1484 pmap_insert(env->regs, (void *) sp, NULL);
1485 pmap_insert(env->regs, (void *) arch_env->bp, NULL);
1486 start_bl = get_irg_start_block(irg);
1487 ir_node *const start = be_new_Start(NULL, start_bl, pmap_count(env->regs) + 1);
1488 set_irg_start(irg, start);
1491 * make proj nodes for the callee save registers.
1492 * memorize them, since Return nodes get those as inputs.
1494 * Note, that if a register corresponds to an argument, the regs map
1495 * contains the old Proj from start for that argument.
1497 rm = ALLOCAN(reg_node_map_t, pmap_count(env->regs));
1498 reg_map_to_arr(rm, env->regs);
1499 for (i = 0, n = pmap_count(env->regs); i < n; ++i) {
1500 const arch_register_t *reg = rm[i].reg;
1501 ir_mode *mode = reg->reg_class->mode;
1503 arch_register_req_type_t add_type = arch_register_req_type_none;
1507 add_type |= arch_register_req_type_produces_sp;
1508 if (!rbitset_is_set(birg->allocatable_regs, reg->global_index)) {
1509 add_type |= arch_register_req_type_ignore;
1513 proj = new_r_Proj(start, mode, nr + 1);
1514 pmap_insert(env->regs, (void *) reg, proj);
1515 be_set_constr_single_reg_out(start, nr + 1, reg, add_type);
1516 arch_set_irn_register(proj, reg);
1518 DBG((dbg, LEVEL_2, "\tregister save proj #%d -> reg %s\n", nr, reg->name));
1521 /* create a new initial memory proj */
1522 assert(is_Proj(old_mem));
1523 arch_set_irn_register_req_out(start, 0, arch_no_register_req);
1524 new_mem_proj = new_r_Proj(start, mode_M, 0);
1526 set_irg_initial_mem(irg, mem);
1528 env->init_sp = be_abi_reg_map_get(env->regs, sp);
1530 /* set new frame_pointer */
1531 frame_pointer = be_abi_reg_map_get(env->regs, fp_reg);
1532 set_irg_frame(irg, frame_pointer);
1534 /* rewire old mem users to new mem */
1535 exchange(old_mem, mem);
1537 /* keep the mem (for functions with an endless loop = no return) */
1540 set_irg_initial_mem(irg, mem);
1542 /* Now, introduce stack param nodes for all parameters passed on the stack */
1543 for (i = 0; i < n_params; ++i) {
1544 ir_node *arg_proj = args[i];
1545 ir_node *repl = NULL;
1547 if (arg_proj != NULL) {
1548 be_abi_call_arg_t *arg;
1549 ir_type *param_type;
1550 int nr = get_Proj_proj(arg_proj);
1553 nr = MIN(nr, n_params);
1554 arg = get_call_arg(call, 0, nr, 1);
1555 param_type = get_method_param_type(method_type, nr);
1558 repl = pmap_get(ir_node, env->regs, arg->reg);
1560 ir_node *addr = be_new_FrameAddr(sp->reg_class, start_bl, frame_pointer, arg->stack_ent);
1562 /* For atomic parameters which are actually used, we create a Load node. */
1563 if (is_atomic_type(param_type) && get_irn_n_edges(args[i]) > 0) {
1564 ir_mode *mode = get_type_mode(param_type);
1565 ir_mode *load_mode = arg->load_mode;
1566 ir_node *nomem = get_irg_no_mem(irg);
1568 ir_node *load = new_r_Load(start_bl, nomem, addr, load_mode, cons_floats);
1569 repl = new_r_Proj(load, load_mode, pn_Load_res);
1571 if (mode != load_mode) {
1572 repl = new_r_Conv(start_bl, repl, mode);
1575 /* The stack parameter is not primitive (it is a struct or array),
1576 * we thus will create a node representing the parameter's address
1582 assert(repl != NULL);
1584 /* Beware: the mode of the register parameters is always the mode of the register class
1585 which may be wrong. Add Conv's then. */
1586 mode = get_irn_mode(args[i]);
1587 if (mode != get_irn_mode(repl)) {
1588 repl = new_r_Conv(get_nodes_block(repl), repl, mode);
1590 exchange(args[i], repl);
1594 /* the arg proj is not needed anymore now and should be only used by the anchor */
1595 assert(get_irn_n_edges(arg_tuple) == 1);
1596 kill_node(arg_tuple);
1597 set_irg_args(irg, new_r_Bad(irg, mode_T));
1599 /* All Return nodes hang on the End node, so look for them there. */
1600 end = get_irg_end_block(irg);
1601 for (i = 0, n = get_Block_n_cfgpreds(end); i < n; ++i) {
1602 ir_node *irn = get_Block_cfgpred(end, i);
1604 if (is_Return(irn)) {
1605 ir_node *blk = get_nodes_block(irn);
1606 ir_node *mem = get_Return_mem(irn);
1607 ir_node *ret = create_be_return(env, irn, blk, mem, get_Return_n_ress(irn));
1612 /* if we have endless loops here, n might be <= 0. Do NOT create a be_Return then,
1613 the code is dead and will never be executed. */
1616 /** Fix the state inputs of calls that still hang on unknowns */
1617 static void fix_call_state_inputs(ir_graph *const irg, be_abi_irg_t *const env)
1619 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1621 arch_register_t **stateregs = NEW_ARR_F(arch_register_t*, 0);
1623 /* Collect caller save registers */
1624 n = arch_env->n_register_classes;
1625 for (i = 0; i < n; ++i) {
1627 const arch_register_class_t *cls = &arch_env->register_classes[i];
1628 for (j = 0; j < cls->n_regs; ++j) {
1629 const arch_register_t *reg = arch_register_for_index(cls, j);
1630 if (reg->type & arch_register_type_state) {
1631 ARR_APP1(arch_register_t*, stateregs, (arch_register_t *)reg);
1636 n = ARR_LEN(env->calls);
1637 n_states = ARR_LEN(stateregs);
1638 for (i = 0; i < n; ++i) {
1640 ir_node *call = env->calls[i];
1642 arity = get_irn_arity(call);
1644 /* the state reg inputs are the last n inputs of the calls */
1645 for (s = 0; s < n_states; ++s) {
1646 int inp = arity - n_states + s;
1647 const arch_register_t *reg = stateregs[s];
1648 ir_node *regnode = be_abi_reg_map_get(env->regs, reg);
1650 set_irn_n(call, inp, regnode);
1654 DEL_ARR_F(stateregs);
1658 * Create a trampoline entity for the given method.
1660 static ir_entity *create_trampoline(be_main_env_t *be, ir_entity *method)
1662 ir_type *type = get_entity_type(method);
1663 ident *old_id = get_entity_ld_ident(method);
1664 ident *id = id_mangle3("", old_id, "$stub");
1665 ir_type *parent = be->pic_trampolines_type;
1666 ir_entity *ent = new_entity(parent, old_id, type);
1667 set_entity_ld_ident(ent, id);
1668 set_entity_visibility(ent, ir_visibility_private);
1674 * Returns the trampoline entity for the given method.
1676 static ir_entity *get_trampoline(be_main_env_t *env, ir_entity *method)
1678 ir_entity *result = pmap_get(ir_entity, env->ent_trampoline_map, method);
1679 if (result == NULL) {
1680 result = create_trampoline(env, method);
1681 pmap_insert(env->ent_trampoline_map, method, result);
1687 static ir_entity *create_pic_symbol(be_main_env_t *be, ir_entity *entity)
1689 ident *old_id = get_entity_ld_ident(entity);
1690 ident *id = id_mangle3("", old_id, "$non_lazy_ptr");
1691 ir_type *e_type = get_entity_type(entity);
1692 ir_type *type = new_type_pointer(e_type);
1693 ir_type *parent = be->pic_symbols_type;
1694 ir_entity *ent = new_entity(parent, old_id, type);
1695 set_entity_ld_ident(ent, id);
1696 set_entity_visibility(ent, ir_visibility_private);
1701 static ir_entity *get_pic_symbol(be_main_env_t *env, ir_entity *entity)
1703 ir_entity *result = pmap_get(ir_entity, env->ent_pic_symbol_map, entity);
1704 if (result == NULL) {
1705 result = create_pic_symbol(env, entity);
1706 pmap_insert(env->ent_pic_symbol_map, entity, result);
1715 * Returns non-zero if a given entity can be accessed using a relative address.
1717 static int can_address_relative(ir_entity *entity)
1719 return entity_has_definition(entity) && !(get_entity_linkage(entity) & IR_LINKAGE_MERGE);
1722 static ir_node *get_pic_base(ir_graph *irg)
1724 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1725 if (arch_env->impl->get_pic_base == NULL)
1727 return arch_env->impl->get_pic_base(irg);
1730 /** patches SymConsts to work in position independent code */
1731 static void fix_pic_symconsts(ir_node *node, void *data)
1733 ir_graph *irg = get_irn_irg(node);
1734 be_main_env_t *be = be_get_irg_main_env(irg);
1744 arity = get_irn_arity(node);
1745 for (i = 0; i < arity; ++i) {
1747 ir_node *pred = get_irn_n(node, i);
1749 ir_entity *pic_symbol;
1750 ir_node *pic_symconst;
1752 if (!is_SymConst(pred))
1755 entity = get_SymConst_entity(pred);
1756 block = get_nodes_block(pred);
1758 /* calls can jump to relative addresses, so we can directly jump to
1759 the (relatively) known call address or the trampoline */
1760 if (i == 1 && is_Call(node)) {
1761 ir_entity *trampoline;
1762 ir_node *trampoline_const;
1764 if (can_address_relative(entity))
1767 dbgi = get_irn_dbg_info(pred);
1768 trampoline = get_trampoline(be, entity);
1769 trampoline_const = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
1771 set_irn_n(node, i, trampoline_const);
1775 /* everything else is accessed relative to EIP */
1776 mode = get_irn_mode(pred);
1777 pic_base = get_pic_base(irg);
1779 /* all ok now for locally constructed stuff */
1780 if (can_address_relative(entity)) {
1781 ir_node *add = new_r_Add(block, pic_base, pred, mode);
1783 /* make sure the walker doesn't visit this add again */
1784 mark_irn_visited(add);
1785 set_irn_n(node, i, add);
1789 /* get entry from pic symbol segment */
1790 dbgi = get_irn_dbg_info(pred);
1791 pic_symbol = get_pic_symbol(be, entity);
1792 pic_symconst = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
1794 add = new_r_Add(block, pic_base, pic_symconst, mode);
1795 mark_irn_visited(add);
1797 /* we need an extra indirection for global data outside our current
1798 module. The loads are always safe and can therefore float
1799 and need no memory input */
1800 load = new_r_Load(block, get_irg_no_mem(irg), add, mode, cons_floats);
1801 load_res = new_r_Proj(load, mode, pn_Load_res);
1803 set_irn_n(node, i, load_res);
1807 void be_abi_introduce(ir_graph *irg)
1809 ir_node *old_frame = get_irg_frame(irg);
1810 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1811 ir_entity *entity = get_irg_entity(irg);
1812 ir_type *method_type = get_entity_type(entity);
1813 be_irg_t *birg = be_birg_from_irg(irg);
1814 struct obstack *obst = &birg->obst;
1815 ir_node *dummy = new_r_Dummy(irg,
1816 arch_env->sp->reg_class->mode);
1819 /* determine allocatable registers */
1820 assert(birg->allocatable_regs == NULL);
1821 birg->allocatable_regs = rbitset_obstack_alloc(obst, arch_env->n_registers);
1822 for (r = 0; r < arch_env->n_registers; ++r) {
1823 const arch_register_t *reg = &arch_env->registers[r];
1824 if ( !(reg->type & arch_register_type_ignore)) {
1825 rbitset_set(birg->allocatable_regs, r);
1829 /* Break here if backend provides a custom API. */
1831 be_omit_fp = be_options.omit_fp;
1834 env.keep_map = pmap_create();
1835 env.call = be_abi_call_new(arch_env->sp->reg_class);
1836 arch_env_get_call_abi(arch_env, method_type, env.call);
1838 env.init_sp = dummy;
1839 env.calls = NEW_ARR_F(ir_node*, 0);
1843 if (be_options.pic) {
1844 irg_walk_graph(irg, fix_pic_symconsts, NULL, NULL);
1847 /* Lower all call nodes in the IRG. */
1848 process_calls(irg, &env);
1850 /* Process the IRG */
1851 modify_irg(irg, &env);
1853 /* fix call inputs for state registers */
1854 fix_call_state_inputs(irg, &env);
1856 be_abi_call_free(env.call);
1858 /* We don't need the keep map anymore. */
1859 pmap_destroy(env.keep_map);
1861 /* calls array is not needed anymore */
1862 DEL_ARR_F(env.calls);
1864 /* reroute the stack origin of the calls to the true stack origin. */
1865 exchange(dummy, env.init_sp);
1866 exchange(old_frame, get_irg_frame(irg));
1868 pmap_destroy(env.regs);
1871 void be_put_allocatable_regs(const ir_graph *irg,
1872 const arch_register_class_t *cls, bitset_t *bs)
1874 be_irg_t *birg = be_birg_from_irg(irg);
1875 unsigned *allocatable_regs = birg->allocatable_regs;
1878 assert(bitset_size(bs) == cls->n_regs);
1879 bitset_clear_all(bs);
1880 for (i = 0; i < cls->n_regs; ++i) {
1881 const arch_register_t *reg = &cls->regs[i];
1882 if (rbitset_is_set(allocatable_regs, reg->global_index))
1887 unsigned be_get_n_allocatable_regs(const ir_graph *irg,
1888 const arch_register_class_t *cls)
1890 bitset_t *bs = bitset_alloca(cls->n_regs);
1891 be_put_allocatable_regs(irg, cls, bs);
1892 return bitset_popcount(bs);
1895 void be_set_allocatable_regs(const ir_graph *irg,
1896 const arch_register_class_t *cls,
1897 unsigned *raw_bitset)
1899 be_irg_t *birg = be_birg_from_irg(irg);
1900 unsigned *allocatable_regs = birg->allocatable_regs;
1903 rbitset_clear_all(raw_bitset, cls->n_regs);
1904 for (i = 0; i < cls->n_regs; ++i) {
1905 const arch_register_t *reg = &cls->regs[i];
1906 if (rbitset_is_set(allocatable_regs, reg->global_index))
1907 rbitset_set(raw_bitset, i);
1911 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_abi)
1912 void be_init_abi(void)
1914 FIRM_DBG_REGISTER(dbg, "firm.be.abi");