2 * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Backend ABI implementation.
23 * @author Sebastian Hack, Michael Beck
32 #include "irgraph_t.h"
35 #include "iredges_t.h"
38 #include "irprintf_t.h"
45 #include "raw_bitset.h"
56 #include "bessaconstr.h"
58 #include "betranshlp.h"
60 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
62 typedef struct be_abi_call_arg_t {
63 unsigned is_res : 1; /**< 1: the call argument is a return value. 0: it's a call parameter. */
64 unsigned in_reg : 1; /**< 1: this argument is transmitted in registers. */
65 unsigned on_stack : 1; /**< 1: this argument is transmitted on the stack. */
66 unsigned callee : 1; /**< 1: someone called us. 0: We call another function */
69 const arch_register_t *reg;
72 unsigned alignment; /**< stack alignment */
73 unsigned space_before; /**< allocate space before */
74 unsigned space_after; /**< allocate space after */
77 struct be_abi_call_t {
78 be_abi_call_flags_t flags; /**< Flags describing the ABI behavior on calls */
79 int pop; /**< number of bytes the stack frame is shrinked by the callee on return. */
80 const be_abi_callbacks_t *cb;
81 ir_type *between_type;
83 const arch_register_class_t *cls_addr; /**< register class of the call address */
87 * The ABI information for the current graph.
90 be_abi_call_t *call; /**< The ABI call information. */
92 ir_node *init_sp; /**< The node representing the stack pointer
93 at the start of the function. */
95 ir_node *start; /**< The be_Start params node. */
96 pmap *regs; /**< A map of all callee-save and ignore regs to
97 their Projs to the RegParams node. */
98 int start_block_bias; /**< The stack bias at the end of the start block. */
100 pmap *keep_map; /**< mapping blocks to keep nodes. */
102 ir_node **calls; /**< flexible array containing all be_Call nodes */
105 static ir_heights_t *ir_heights;
107 /** Flag: if set, try to omit the frame pointer in all routines. */
108 static int be_omit_fp = 1;
110 static ir_node *be_abi_reg_map_get(pmap *map, const arch_register_t *reg)
112 return (ir_node*)pmap_get(map, reg);
115 static void be_abi_reg_map_set(pmap *map, const arch_register_t* reg,
118 pmap_insert(map, reg, node);
122 _ ____ ___ ____ _ _ _ _
123 / \ | __ )_ _| / ___|__ _| | | |__ __ _ ___| | _____
124 / _ \ | _ \| | | | / _` | | | '_ \ / _` |/ __| |/ / __|
125 / ___ \| |_) | | | |__| (_| | | | |_) | (_| | (__| <\__ \
126 /_/ \_\____/___| \____\__,_|_|_|_.__/ \__,_|\___|_|\_\___/
128 These callbacks are used by the backend to set the parameters
129 for a specific call type.
133 * Set compare function: compares two ABI call object arguments.
135 static int cmp_call_arg(const void *a, const void *b, size_t n)
137 const be_abi_call_arg_t *p = (const be_abi_call_arg_t*)a;
138 const be_abi_call_arg_t *q = (const be_abi_call_arg_t*)b;
140 return !(p->is_res == q->is_res && p->pos == q->pos && p->callee == q->callee);
144 * Get an ABI call object argument.
146 * @param call the abi call
147 * @param is_res true for call results, false for call arguments
148 * @param pos position of the argument
149 * @param callee context type - if we are callee or caller
151 static be_abi_call_arg_t *get_call_arg(be_abi_call_t *call, int is_res, int pos, int callee)
153 be_abi_call_arg_t arg;
156 memset(&arg, 0, sizeof(arg));
161 hash = is_res * 128 + pos;
163 return (be_abi_call_arg_t*)set_find(call->params, &arg, sizeof(arg), hash);
167 * Set an ABI call object argument.
169 static void remember_call_arg(be_abi_call_arg_t *arg, be_abi_call_t *call, be_abi_context_t context)
171 unsigned hash = arg->is_res * 128 + arg->pos;
172 if (context & ABI_CONTEXT_CALLEE) {
174 set_insert(call->params, arg, sizeof(*arg), hash);
176 if (context & ABI_CONTEXT_CALLER) {
178 set_insert(call->params, arg, sizeof(*arg), hash);
182 /* Set the flags for a call. */
183 void be_abi_call_set_flags(be_abi_call_t *call, be_abi_call_flags_t flags, const be_abi_callbacks_t *cb)
189 /* Sets the number of bytes the stackframe is shrinked by the callee on return */
190 void be_abi_call_set_pop(be_abi_call_t *call, int pop)
196 /* Set register class for call address */
197 void be_abi_call_set_call_address_reg_class(be_abi_call_t *call, const arch_register_class_t *cls)
199 call->cls_addr = cls;
203 void be_abi_call_param_stack(be_abi_call_t *call, int arg_pos,
204 ir_mode *load_mode, unsigned alignment,
205 unsigned space_before, unsigned space_after,
206 be_abi_context_t context)
208 be_abi_call_arg_t arg;
209 memset(&arg, 0, sizeof(arg));
210 assert(alignment > 0 && "Alignment must be greater than 0");
212 arg.load_mode = load_mode;
213 arg.alignment = alignment;
214 arg.space_before = space_before;
215 arg.space_after = space_after;
219 remember_call_arg(&arg, call, context);
222 void be_abi_call_param_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
224 be_abi_call_arg_t arg;
225 memset(&arg, 0, sizeof(arg));
232 remember_call_arg(&arg, call, context);
235 void be_abi_call_res_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
237 be_abi_call_arg_t arg;
238 memset(&arg, 0, sizeof(arg));
245 remember_call_arg(&arg, call, context);
248 /* Get the flags of a ABI call object. */
249 be_abi_call_flags_t be_abi_call_get_flags(const be_abi_call_t *call)
255 * Constructor for a new ABI call object.
257 * @param cls_addr register class of the call address
259 * @return the new ABI call object
261 static be_abi_call_t *be_abi_call_new(const arch_register_class_t *cls_addr)
263 be_abi_call_t *call = XMALLOCZ(be_abi_call_t);
266 call->params = new_set(cmp_call_arg, 16);
268 call->cls_addr = cls_addr;
270 call->flags.bits.try_omit_fp = be_omit_fp;
276 * Destructor for an ABI call object.
278 static void be_abi_call_free(be_abi_call_t *call)
280 del_set(call->params);
285 * Initializes the frame layout from parts
287 * @param frame the stack layout that will be initialized
288 * @param args the stack argument layout type
289 * @param between the between layout type
290 * @param locals the method frame type
291 * @param param_map an array mapping method argument positions to the stack argument type
293 * @return the initialized stack layout
295 static be_stack_layout_t *stack_frame_init(be_stack_layout_t *frame, ir_type *args,
296 ir_type *between, ir_type *locals,
297 ir_entity *param_map[])
299 frame->arg_type = args;
300 frame->between_type = between;
301 frame->frame_type = locals;
302 frame->initial_offset = 0;
303 frame->initial_bias = 0;
304 frame->order[1] = between;
305 frame->param_map = param_map;
307 /* typical decreasing stack: locals have the
308 * lowest addresses, arguments the highest */
309 frame->order[0] = locals;
310 frame->order[2] = args;
321 Adjustment of the calls inside a graph.
326 * Transform a call node into a be_Call node.
328 * @param env The ABI environment for the current irg.
329 * @param irn The call node.
330 * @param curr_sp The stack pointer node to use.
331 * @return The stack pointer after the call.
333 static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
335 ir_graph *irg = get_irn_irg(irn);
336 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
337 ir_type *call_tp = get_Call_type(irn);
338 ir_node *call_ptr = get_Call_ptr(irn);
339 size_t n_params = get_method_n_params(call_tp);
340 ir_node *curr_mem = get_Call_mem(irn);
341 ir_node *bl = get_nodes_block(irn);
343 const arch_register_t *sp = arch_env->sp;
344 be_abi_call_t *call = be_abi_call_new(sp->reg_class);
345 ir_mode *mach_mode = sp->reg_class->mode;
346 int no_alloc = call->flags.bits.frame_is_setup_on_call;
347 int n_res = get_method_n_ress(call_tp);
348 int do_seq = call->flags.bits.store_args_sequential && !no_alloc;
350 ir_node *res_proj = NULL;
351 int n_reg_params = 0;
352 int n_stack_params = 0;
355 const arch_register_t **states = NEW_ARR_F(const arch_register_t*, 0);
356 const arch_register_t **destroyed_regs = NEW_ARR_F(const arch_register_t*, 0);
360 int n_reg_results = 0;
361 const ir_edge_t *edge;
363 int *stack_param_idx;
364 int i, n, destroy_all_regs;
365 int throws_exception;
370 /* Let the isa fill out the abi description for that call node. */
371 arch_env_get_call_abi(arch_env, call_tp, call);
373 /* Insert code to put the stack arguments on the stack. */
374 assert(get_Call_n_params(irn) == n_params);
375 stack_param_idx = ALLOCAN(int, n_params);
376 for (p = 0; p < n_params; ++p) {
377 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
380 int arg_size = get_type_size_bytes(get_method_param_type(call_tp, p));
382 stack_size += round_up2(arg->space_before, arg->alignment);
383 stack_size += round_up2(arg_size, arg->alignment);
384 stack_size += round_up2(arg->space_after, arg->alignment);
386 stack_param_idx[n_stack_params++] = p;
390 /* Collect all arguments which are passed in registers. */
391 reg_param_idxs = 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);
394 if (arg && arg->in_reg) {
395 reg_param_idxs[n_reg_params++] = p;
400 * If the stack is decreasing and we do not want to store sequentially,
401 * or someone else allocated the call frame
402 * we allocate as much space on the stack all parameters need, by
403 * moving the stack pointer along the stack's direction.
405 * Note: we also have to do this for stack_size == 0, because we may have
406 * to adjust stack alignment for the call.
408 if (!do_seq && !no_alloc) {
409 curr_sp = be_new_IncSP(sp, bl, curr_sp, stack_size, 1);
412 dbgi = get_irn_dbg_info(irn);
413 /* If there are some parameters which shall be passed on the stack. */
414 if (n_stack_params > 0) {
416 ir_node **in = ALLOCAN(ir_node*, n_stack_params+1);
419 /* push params in reverse direction because stack grows downwards */
421 for (i = 0; i < n_stack_params >> 1; ++i) {
422 int other = n_stack_params - i - 1;
423 int tmp = stack_param_idx[i];
424 stack_param_idx[i] = stack_param_idx[other];
425 stack_param_idx[other] = tmp;
429 curr_mem = get_Call_mem(irn);
431 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.
450 addr = curr_sp = be_new_IncSP(sp, bl, curr_sp,
451 param_size + arg->space_before, 0);
452 add_irn_dep(curr_sp, curr_mem);
454 curr_ofs += arg->space_before;
455 curr_ofs = round_up2(curr_ofs, arg->alignment);
457 /* Make the expression to compute the argument's offset. */
459 ir_mode *constmode = mach_mode;
460 if (mode_is_reference(mach_mode)) {
463 addr = new_r_Const_long(irg, constmode, curr_ofs);
464 addr = new_r_Add(bl, curr_sp, addr, mach_mode);
468 /* Insert a store for primitive arguments. */
469 if (is_atomic_type(param_type)) {
470 ir_node *nomem = get_irg_no_mem(irg);
471 ir_node *mem_input = do_seq ? curr_mem : nomem;
472 ir_node *store = new_rd_Store(dbgi, bl, mem_input, addr, param, cons_none);
473 mem = new_r_Proj(store, mode_M, pn_Store_M);
475 /* Make a mem copy for compound arguments. */
478 assert(mode_is_reference(get_irn_mode(param)));
479 copy = new_rd_CopyB(dbgi, bl, curr_mem, addr, param, param_type);
480 mem = new_r_Proj(copy, mode_M, pn_CopyB_M);
483 curr_ofs += param_size;
491 /* We need the sync only, if we didn't build the stores sequentially. */
493 if (n_stack_params >= 1) {
494 curr_mem = new_r_Sync(bl, n_in, in);
496 curr_mem = get_Call_mem(irn);
501 /* check for the return_twice property */
502 destroy_all_regs = 0;
503 if (is_SymConst_addr_ent(call_ptr)) {
504 ir_entity *ent = get_SymConst_entity(call_ptr);
506 if (get_entity_additional_properties(ent) & mtp_property_returns_twice)
507 destroy_all_regs = 1;
509 ir_type *call_tp = get_Call_type(irn);
511 if (get_method_additional_properties(call_tp) & mtp_property_returns_twice)
512 destroy_all_regs = 1;
515 /* Put caller save into the destroyed set and state registers in the states
517 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
519 const arch_register_class_t *cls = &arch_env->register_classes[i];
520 for (j = 0; j < cls->n_regs; ++j) {
521 const arch_register_t *reg = arch_register_for_index(cls, j);
523 /* even if destroyed all is specified, neither SP nor FP are
524 * destroyed (else bad things will happen) */
525 if (reg == arch_env->sp || reg == arch_env->bp)
528 if (reg->type & arch_register_type_state) {
529 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
530 ARR_APP1(const arch_register_t*, states, reg);
531 /* we're already in the destroyed set so no need for further
535 if (destroy_all_regs || arch_register_is_caller_save(arch_env, reg)) {
536 if (!(reg->type & arch_register_type_ignore)) {
537 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
543 /* search the largest result proj number */
544 res_projs = ALLOCANZ(ir_node*, n_res);
546 foreach_out_edge(irn, edge) {
547 const ir_edge_t *res_edge;
548 ir_node *irn = get_edge_src_irn(edge);
550 if (!is_Proj(irn) || get_Proj_proj(irn) != pn_Call_T_result)
553 foreach_out_edge(irn, res_edge) {
555 ir_node *res = get_edge_src_irn(res_edge);
557 assert(is_Proj(res));
559 proj = get_Proj_proj(res);
560 assert(proj < n_res);
561 assert(res_projs[proj] == NULL);
562 res_projs[proj] = res;
568 /** TODO: this is not correct for cases where return values are passed
569 * on the stack, but no known ABI does this currently...
571 n_reg_results = n_res;
574 in = ALLOCAN(ir_node*, n_reg_params + ARR_LEN(states));
576 /* make the back end call node and set its register requirements. */
577 for (i = 0; i < n_reg_params; ++i) {
578 in[n_ins++] = get_Call_param(irn, reg_param_idxs[i]);
581 /* add state registers ins */
582 for (s = 0; s < ARR_LEN(states); ++s) {
583 const arch_register_t *reg = states[s];
584 const arch_register_class_t *cls = arch_register_get_class(reg);
585 ir_node *regnode = new_r_Unknown(irg, arch_register_class_mode(cls));
586 in[n_ins++] = regnode;
588 assert(n_ins == (int) (n_reg_params + ARR_LEN(states)));
590 /* ins collected, build the call */
591 throws_exception = ir_throws_exception(irn);
592 if (env->call->flags.bits.call_has_imm && is_SymConst(call_ptr)) {
594 low_call = be_new_Call(dbgi, irg, bl, curr_mem, curr_sp, curr_sp,
595 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
596 n_ins, in, get_Call_type(irn));
597 be_Call_set_entity(low_call, get_SymConst_entity(call_ptr));
600 low_call = be_new_Call(dbgi, irg, bl, curr_mem, curr_sp, call_ptr,
601 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
602 n_ins, in, get_Call_type(irn));
604 ir_set_throws_exception(low_call, throws_exception);
605 be_Call_set_pop(low_call, call->pop);
607 /* put the call into the list of all calls for later processing */
608 ARR_APP1(ir_node *, env->calls, low_call);
610 /* create new stack pointer */
611 curr_sp = new_r_Proj(low_call, get_irn_mode(curr_sp), pn_be_Call_sp);
612 be_set_constr_single_reg_out(low_call, pn_be_Call_sp, sp,
613 arch_register_req_type_ignore | arch_register_req_type_produces_sp);
614 arch_set_irn_register(curr_sp, sp);
616 /* now handle results */
617 for (i = 0; i < n_res; ++i) {
618 ir_node *proj = res_projs[i];
619 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
620 long pn = i + pn_be_Call_first_res;
622 /* returns values on stack not supported yet */
626 shift the proj number to the right, since we will drop the
627 unspeakable Proj_T from the Call. Therefore, all real argument
628 Proj numbers must be increased by pn_be_Call_first_res
630 pn = i + pn_be_Call_first_res;
633 ir_type *res_type = get_method_res_type(call_tp, i);
634 ir_mode *mode = get_type_mode(res_type);
635 proj = new_r_Proj(low_call, mode, pn);
638 set_Proj_pred(proj, low_call);
639 set_Proj_proj(proj, pn);
643 /* remove register from destroyed regs */
645 size_t n = ARR_LEN(destroyed_regs);
646 for (j = 0; j < n; ++j) {
647 if (destroyed_regs[j] == arg->reg) {
648 destroyed_regs[j] = destroyed_regs[n-1];
649 ARR_SHRINKLEN(destroyed_regs,n-1);
657 Set the register class of the call address to
658 the backend provided class (default: stack pointer class)
660 be_node_set_reg_class_in(low_call, n_be_Call_ptr, call->cls_addr);
662 DBG((dbg, LEVEL_3, "\tcreated backend call %+F\n", low_call));
664 /* Set the register classes and constraints of the Call parameters. */
665 for (i = 0; i < n_reg_params; ++i) {
666 int index = reg_param_idxs[i];
667 be_abi_call_arg_t *arg = get_call_arg(call, 0, index, 0);
668 assert(arg->reg != NULL);
670 be_set_constr_single_reg_in(low_call, n_be_Call_first_arg + i,
671 arg->reg, arch_register_req_type_none);
674 /* Set the register constraints of the results. */
675 for (i = 0; i < n_res; ++i) {
676 ir_node *proj = res_projs[i];
677 const be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
678 int pn = get_Proj_proj(proj);
681 be_set_constr_single_reg_out(low_call, pn, arg->reg,
682 arch_register_req_type_none);
683 arch_set_irn_register(proj, arg->reg);
685 exchange(irn, low_call);
687 /* kill the ProjT node */
688 if (res_proj != NULL) {
692 /* Make additional projs for the caller save registers
693 and the Keep node which keeps them alive. */
699 int curr_res_proj = pn_be_Call_first_res + n_reg_results;
702 n_ins = ARR_LEN(destroyed_regs) + n_reg_results + 1;
703 in = ALLOCAN(ir_node *, n_ins);
705 /* also keep the stack pointer */
706 set_irn_link(curr_sp, (void*) sp);
709 for (d = 0; d < ARR_LEN(destroyed_regs); ++d) {
710 const arch_register_t *reg = destroyed_regs[d];
711 ir_node *proj = new_r_Proj(low_call, reg->reg_class->mode, curr_res_proj);
713 /* memorize the register in the link field. we need afterwards to set the register class of the keep correctly. */
714 be_set_constr_single_reg_out(low_call, curr_res_proj, reg,
715 arch_register_req_type_none);
716 arch_set_irn_register(proj, reg);
718 set_irn_link(proj, (void*) reg);
723 for (i = 0; i < n_reg_results; ++i) {
724 ir_node *proj = res_projs[i];
725 const arch_register_t *reg = arch_get_irn_register(proj);
726 set_irn_link(proj, (void*) reg);
731 /* create the Keep for the caller save registers */
732 keep = be_new_Keep(bl, n, in);
733 for (i = 0; i < n; ++i) {
734 const arch_register_t *reg = (const arch_register_t*)get_irn_link(in[i]);
735 be_node_set_reg_class_in(keep, i, reg->reg_class);
739 /* Clean up the stack. */
740 assert(stack_size >= call->pop);
741 stack_size -= call->pop;
743 if (stack_size > 0) {
744 ir_node *mem_proj = NULL;
746 foreach_out_edge(low_call, edge) {
747 ir_node *irn = get_edge_src_irn(edge);
748 if (is_Proj(irn) && get_Proj_proj(irn) == pn_Call_M) {
755 mem_proj = new_r_Proj(low_call, mode_M, pn_be_Call_M);
756 keep_alive(mem_proj);
759 /* Clean up the stack frame or revert alignment fixes if we allocated it */
761 curr_sp = be_new_IncSP(sp, bl, curr_sp, -stack_size, 0);
764 be_abi_call_free(call);
767 DEL_ARR_F(destroyed_regs);
773 * Adjust the size of a node representing a stack alloc or free for the minimum stack alignment.
775 * @param alignment the minimum stack alignment
776 * @param size the node containing the non-aligned size
777 * @param block the block where new nodes are allocated on
778 * @param dbg debug info for new nodes
780 * @return a node representing the aligned size
782 static ir_node *adjust_alloc_size(unsigned stack_alignment, ir_node *size,
783 ir_node *block, dbg_info *dbg)
785 if (stack_alignment > 1) {
791 assert(is_po2(stack_alignment));
793 mode = get_irn_mode(size);
794 tv = new_tarval_from_long(stack_alignment-1, mode);
795 irg = get_Block_irg(block);
796 mask = new_r_Const(irg, tv);
797 size = new_rd_Add(dbg, block, size, mask, mode);
799 tv = new_tarval_from_long(-(long)stack_alignment, mode);
800 mask = new_r_Const(irg, tv);
801 size = new_rd_And(dbg, block, size, mask, mode);
807 * The alloca is transformed into a back end alloca node and connected to the stack nodes.
809 static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp)
811 ir_node *block = get_nodes_block(alloc);
812 ir_graph *irg = get_Block_irg(block);
813 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
814 ir_node *alloc_mem = NULL;
815 ir_node *alloc_res = NULL;
816 ir_type *type = get_Alloc_type(alloc);
819 const ir_edge_t *edge;
824 unsigned stack_alignment;
826 /* all non-stack Alloc nodes should already be lowered before the backend */
827 assert(get_Alloc_where(alloc) == stack_alloc);
829 foreach_out_edge(alloc, edge) {
830 ir_node *irn = get_edge_src_irn(edge);
832 assert(is_Proj(irn));
833 switch (get_Proj_proj(irn)) {
845 /* Beware: currently Alloc nodes without a result might happen,
846 only escape analysis kills them and this phase runs only for object
847 oriented source. We kill the Alloc here. */
848 if (alloc_res == NULL && alloc_mem) {
849 exchange(alloc_mem, get_Alloc_mem(alloc));
853 dbg = get_irn_dbg_info(alloc);
854 count = get_Alloc_count(alloc);
856 /* we might need to multiply the count with the element size */
857 if (type != firm_unknown_type && get_type_size_bytes(type) != 1) {
858 ir_mode *mode = get_irn_mode(count);
859 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type),
861 ir_node *cnst = new_rd_Const(dbg, irg, tv);
862 size = new_rd_Mul(dbg, block, count, cnst, mode);
867 /* The stack pointer will be modified in an unknown manner.
868 We cannot omit it. */
869 env->call->flags.bits.try_omit_fp = 0;
871 stack_alignment = 1 << arch_env->stack_alignment;
872 size = adjust_alloc_size(stack_alignment, size, block, dbg);
873 new_alloc = be_new_AddSP(arch_env->sp, block, curr_sp, size);
874 set_irn_dbg_info(new_alloc, dbg);
876 if (alloc_mem != NULL) {
880 addsp_mem = new_r_Proj(new_alloc, mode_M, pn_be_AddSP_M);
882 /* We need to sync the output mem of the AddSP with the input mem
883 edge into the alloc node. */
884 ins[0] = get_Alloc_mem(alloc);
886 sync = new_r_Sync(block, 2, ins);
888 exchange(alloc_mem, sync);
891 exchange(alloc, new_alloc);
893 /* fix projnum of alloca res */
894 set_Proj_proj(alloc_res, pn_be_AddSP_res);
896 curr_sp = new_r_Proj(new_alloc, get_irn_mode(curr_sp), pn_be_AddSP_sp);
903 * The Free is transformed into a back end free node and connected to the stack nodes.
905 static ir_node *adjust_free(be_abi_irg_t *env, ir_node *free, ir_node *curr_sp)
907 ir_node *block = get_nodes_block(free);
908 ir_graph *irg = get_irn_irg(free);
909 ir_type *type = get_Free_type(free);
910 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
911 ir_mode *sp_mode = arch_env->sp->reg_class->mode;
912 dbg_info *dbg = get_irn_dbg_info(free);
913 ir_node *subsp, *mem, *res, *size, *sync;
915 unsigned stack_alignment;
917 /* all non-stack-alloc Free nodes should already be lowered before the
919 assert(get_Free_where(free) == stack_alloc);
921 /* we might need to multiply the size with the element size */
922 if (type != firm_unknown_type && get_type_size_bytes(type) != 1) {
923 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type), mode_Iu);
924 ir_node *cnst = new_rd_Const(dbg, irg, tv);
925 ir_node *mul = new_rd_Mul(dbg, block, get_Free_size(free),
929 size = get_Free_size(free);
932 stack_alignment = 1 << arch_env->stack_alignment;
933 size = adjust_alloc_size(stack_alignment, size, block, dbg);
935 /* The stack pointer will be modified in an unknown manner.
936 We cannot omit it. */
937 env->call->flags.bits.try_omit_fp = 0;
938 subsp = be_new_SubSP(arch_env->sp, block, curr_sp, size);
939 set_irn_dbg_info(subsp, dbg);
941 mem = new_r_Proj(subsp, mode_M, pn_be_SubSP_M);
942 res = new_r_Proj(subsp, sp_mode, pn_be_SubSP_sp);
944 /* we need to sync the memory */
945 in[0] = get_Free_mem(free);
947 sync = new_r_Sync(block, 2, in);
949 /* and make the AddSP dependent on the former memory */
950 add_irn_dep(subsp, get_Free_mem(free));
953 exchange(free, sync);
960 * Check if a node is somehow data dependent on another one.
961 * both nodes must be in the same basic block.
962 * @param n1 The first node.
963 * @param n2 The second node.
964 * @return 1, if n1 is data dependent (transitively) on n2, 0 if not.
966 static int dependent_on(ir_node *n1, ir_node *n2)
968 assert(get_nodes_block(n1) == get_nodes_block(n2));
970 return heights_reachable_in_block(ir_heights, n1, n2);
973 static int cmp_call_dependency(const void *c1, const void *c2)
975 ir_node *n1 = *(ir_node **) c1;
976 ir_node *n2 = *(ir_node **) c2;
980 Classical qsort() comparison function behavior:
981 0 if both elements are equal
982 1 if second is "smaller" that first
983 -1 if first is "smaller" that second
985 if (dependent_on(n1, n2))
988 if (dependent_on(n2, n1))
991 /* The nodes have no depth order, but we need a total order because qsort()
994 * Additionally, we need to respect transitive dependencies. Consider a
995 * Call a depending on Call b and an independent Call c.
996 * We MUST NOT order c > a and b > c. */
997 h1 = get_irn_height(ir_heights, n1);
998 h2 = get_irn_height(ir_heights, n2);
999 if (h1 < h2) return -1;
1000 if (h1 > h2) return 1;
1001 /* Same height, so use a random (but stable) order */
1002 return get_irn_idx(n1) - get_irn_idx(n2);
1006 * Walker: links all Call/Alloc/Free nodes to the Block they are contained.
1007 * Clears the irg_is_leaf flag if a Call is detected.
1009 static void link_ops_in_block_walker(ir_node *irn, void *data)
1011 be_abi_irg_t *env = (be_abi_irg_t*)data;
1012 unsigned code = get_irn_opcode(irn);
1014 if (code == iro_Call ||
1015 (code == iro_Alloc && get_Alloc_where(irn) == stack_alloc) ||
1016 (code == iro_Free && get_Free_where(irn) == stack_alloc)) {
1017 ir_node *bl = get_nodes_block(irn);
1018 void *save = get_irn_link(bl);
1020 if (code == iro_Call)
1021 env->call->flags.bits.irg_is_leaf = 0;
1023 set_irn_link(irn, save);
1024 set_irn_link(bl, irn);
1027 if (code == iro_Builtin && get_Builtin_kind(irn) == ir_bk_return_address) {
1028 ir_node *param = get_Builtin_param(irn, 0);
1029 ir_tarval *tv = get_Const_tarval(param);
1030 unsigned long value = get_tarval_long(tv);
1031 /* use ebp, so the climbframe algo works... */
1033 env->call->flags.bits.try_omit_fp = 0;
1040 * Process all Call/Alloc/Free nodes inside a basic block.
1041 * Note that the link field of the block must contain a linked list of all
1042 * nodes inside the Block. We first order this list according to data dependency
1043 * and that connect the nodes together.
1045 static void process_ops_in_block(ir_node *bl, void *data)
1047 be_abi_irg_t *env = (be_abi_irg_t*)data;
1048 ir_node *curr_sp = env->init_sp;
1055 for (irn = (ir_node*)get_irn_link(bl); irn != NULL;
1056 irn = (ir_node*)get_irn_link(irn)) {
1060 nodes = ALLOCAN(ir_node*, n_nodes);
1061 for (irn = (ir_node*)get_irn_link(bl), n = 0; irn != NULL;
1062 irn = (ir_node*)get_irn_link(irn), ++n) {
1066 /* If there were call nodes in the block. */
1071 /* order the call nodes according to data dependency */
1072 qsort(nodes, n_nodes, sizeof(nodes[0]), cmp_call_dependency);
1074 for (i = n_nodes - 1; i >= 0; --i) {
1075 ir_node *irn = nodes[i];
1077 DBG((dbg, LEVEL_3, "\tprocessing call %+F\n", irn));
1078 switch (get_irn_opcode(irn)) {
1081 /* The stack pointer will be modified due to a call. */
1082 env->call->flags.bits.try_omit_fp = 0;
1084 curr_sp = adjust_call(env, irn, curr_sp);
1087 if (get_Alloc_where(irn) == stack_alloc)
1088 curr_sp = adjust_alloc(env, irn, curr_sp);
1091 if (get_Free_where(irn) == stack_alloc)
1092 curr_sp = adjust_free(env, irn, curr_sp);
1095 panic("invalid call");
1099 /* Keep the last stack state in the block by tying it to Keep node,
1100 * the proj from calls is already kept */
1101 if (curr_sp != env->init_sp &&
1102 !(is_Proj(curr_sp) && be_is_Call(get_Proj_pred(curr_sp)))) {
1104 keep = be_new_Keep(bl, 1, nodes);
1105 pmap_insert(env->keep_map, bl, keep);
1109 set_irn_link(bl, curr_sp);
1113 * Adjust all call nodes in the graph to the ABI conventions.
1115 static void process_calls(ir_graph *irg)
1117 be_abi_irg_t *abi = be_get_irg_abi(irg);
1119 abi->call->flags.bits.irg_is_leaf = 1;
1120 irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, abi);
1122 ir_heights = heights_new(irg);
1123 irg_block_walk_graph(irg, NULL, process_ops_in_block, abi);
1124 heights_free(ir_heights);
1128 * Computes the stack argument layout type.
1129 * Changes a possibly allocated value param type by moving
1130 * entities to the stack layout type.
1132 * @param call the current call ABI
1133 * @param method_type the method type
1134 * @param val_param_tp the value parameter type, will be destroyed
1135 * @param param_map an array mapping method arguments to the stack layout type
1137 * @return the stack argument layout type
1139 static ir_type *compute_arg_type(ir_graph *irg, be_abi_call_t *call,
1140 ir_type *method_type, ir_type *val_param_tp,
1141 ir_entity ***param_map)
1143 int n = get_method_n_params(method_type);
1144 struct obstack *obst = be_get_be_obst(irg);
1150 ident *id = get_entity_ident(get_irg_entity(irg));
1153 *param_map = map = OALLOCN(obst, ir_entity*, n);
1154 res = new_type_struct(id_mangle_u(id, new_id_from_chars("arg_type", 8)));
1155 for (i = 0; i < n; ++i) {
1156 ir_type *param_type = get_method_param_type(method_type, i);
1157 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1160 if (arg->on_stack) {
1161 if (val_param_tp != NULL) {
1162 /* the entity was already created, create a copy in the param type */
1163 ir_entity *val_ent = get_method_value_param_ent(method_type, i);
1164 arg->stack_ent = copy_entity_own(val_ent, res);
1165 set_entity_link(val_ent, arg->stack_ent);
1166 set_entity_link(arg->stack_ent, NULL);
1168 /* create a new entity */
1169 snprintf(buf, sizeof(buf), "param_%d", i);
1170 arg->stack_ent = new_entity(res, new_id_from_str(buf), param_type);
1172 ofs += arg->space_before;
1173 ofs = round_up2(ofs, arg->alignment);
1174 set_entity_offset(arg->stack_ent, ofs);
1175 ofs += arg->space_after;
1176 ofs += get_type_size_bytes(param_type);
1177 map[i] = arg->stack_ent;
1180 set_type_size_bytes(res, ofs);
1181 set_type_state(res, layout_fixed);
1186 const arch_register_t *reg;
1190 static int cmp_regs(const void *a, const void *b)
1192 const reg_node_map_t *p = (const reg_node_map_t*)a;
1193 const reg_node_map_t *q = (const reg_node_map_t*)b;
1195 if (p->reg->reg_class == q->reg->reg_class)
1196 return p->reg->index - q->reg->index;
1198 return p->reg->reg_class < q->reg->reg_class ? -1 : +1;
1201 static void reg_map_to_arr(reg_node_map_t *res, pmap *reg_map)
1204 size_t n = pmap_count(reg_map);
1207 foreach_pmap(reg_map, ent) {
1208 res[i].reg = (const arch_register_t*)ent->key;
1209 res[i].irn = (ir_node*)ent->value;
1213 qsort(res, n, sizeof(res[0]), cmp_regs);
1217 * Creates a be_Return for a Return node.
1219 * @param @env the abi environment
1220 * @param irn the Return node or NULL if there was none
1221 * @param bl the block where the be_Retun should be placed
1222 * @param mem the current memory
1223 * @param n_res number of return results
1225 static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl,
1226 ir_node *mem, int n_res)
1228 be_abi_call_t *call = env->call;
1229 ir_graph *irg = get_Block_irg(bl);
1230 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1232 pmap *reg_map = pmap_create();
1233 ir_node *keep = (ir_node*)pmap_get(env->keep_map, bl);
1240 const arch_register_t **regs;
1244 get the valid stack node in this block.
1245 If we had a call in that block there is a Keep constructed by process_calls()
1246 which points to the last stack modification in that block. we'll use
1247 it then. Else we use the stack from the start block and let
1248 the ssa construction fix the usage.
1250 stack = be_abi_reg_map_get(env->regs, arch_env->sp);
1252 stack = get_irn_n(keep, 0);
1254 remove_End_keepalive(get_irg_end(irg), keep);
1257 /* Insert results for Return into the register map. */
1258 for (i = 0; i < n_res; ++i) {
1259 ir_node *res = get_Return_res(irn, i);
1260 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1261 assert(arg->in_reg && "return value must be passed in register");
1262 pmap_insert(reg_map, (void *) arg->reg, res);
1265 /* Add uses of the callee save registers. */
1266 foreach_pmap(env->regs, ent) {
1267 const arch_register_t *reg = (const arch_register_t*)ent->key;
1268 if ((reg->type & arch_register_type_ignore) || arch_register_is_callee_save(arch_env, reg))
1269 pmap_insert(reg_map, ent->key, ent->value);
1272 be_abi_reg_map_set(reg_map, arch_env->sp, stack);
1275 Maximum size of the in array for Return nodes is
1276 return args + callee save/ignore registers + memory + stack pointer
1278 in_max = pmap_count(reg_map) + n_res + 2;
1280 in = ALLOCAN(ir_node*, in_max);
1281 regs = ALLOCAN(arch_register_t const*, in_max);
1284 in[1] = be_abi_reg_map_get(reg_map, arch_env->sp);
1286 regs[1] = arch_env->sp;
1289 /* clear SP entry, since it has already been grown. */
1290 pmap_insert(reg_map, (void *) arch_env->sp, NULL);
1291 for (i = 0; i < n_res; ++i) {
1292 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1294 in[n] = be_abi_reg_map_get(reg_map, arg->reg);
1295 regs[n++] = arg->reg;
1297 /* Clear the map entry to mark the register as processed. */
1298 be_abi_reg_map_set(reg_map, arg->reg, NULL);
1301 /* grow the rest of the stuff. */
1302 foreach_pmap(reg_map, ent) {
1304 in[n] = (ir_node*)ent->value;
1305 regs[n++] = (const arch_register_t*)ent->key;
1309 /* The in array for the new back end return is now ready. */
1311 dbgi = get_irn_dbg_info(irn);
1315 /* we have to pop the shadow parameter in in case of struct returns */
1317 ret = be_new_Return(dbgi, irg, bl, n_res, pop, n, in);
1319 /* Set the register classes of the return's parameter accordingly. */
1320 for (i = 0; i < n; ++i) {
1321 if (regs[i] == NULL)
1324 be_set_constr_single_reg_in(ret, i, regs[i], arch_register_req_type_none);
1327 /* Free the space of the Epilog's in array and the register <-> proj map. */
1328 pmap_destroy(reg_map);
1333 typedef struct ent_pos_pair ent_pos_pair;
1334 struct ent_pos_pair {
1335 ir_entity *ent; /**< a value param entity */
1336 int pos; /**< its parameter number */
1337 ent_pos_pair *next; /**< for linking */
1340 typedef struct lower_frame_sels_env_t {
1341 ent_pos_pair *value_param_list; /**< the list of all value param entities */
1342 ir_node *frame; /**< the current frame */
1343 const arch_register_class_t *sp_class; /**< register class of the stack pointer */
1344 const arch_register_class_t *link_class; /**< register class of the link pointer */
1345 ir_type *value_tp; /**< the value type if any */
1346 ir_type *frame_tp; /**< the frame type */
1347 int static_link_pos; /**< argument number of the hidden static link */
1348 } lower_frame_sels_env_t;
1351 * Return an entity from the backend for an value param entity.
1353 * @param ent an value param type entity
1354 * @param ctx context
1356 static ir_entity *get_argument_entity(ir_entity *ent, lower_frame_sels_env_t *ctx)
1358 ir_entity *argument_ent = (ir_entity*)get_entity_link(ent);
1360 if (argument_ent == NULL) {
1361 /* we have NO argument entity yet: This is bad, as we will
1362 * need one for backing store.
1365 ir_type *frame_tp = ctx->frame_tp;
1366 unsigned offset = get_type_size_bytes(frame_tp);
1367 ir_type *tp = get_entity_type(ent);
1368 unsigned align = get_type_alignment_bytes(tp);
1370 offset += align - 1;
1371 offset &= ~(align - 1);
1373 argument_ent = copy_entity_own(ent, frame_tp);
1375 /* must be automatic to set a fixed layout */
1376 set_entity_offset(argument_ent, offset);
1377 offset += get_type_size_bytes(tp);
1379 set_type_size_bytes(frame_tp, offset);
1380 set_entity_link(ent, argument_ent);
1382 return argument_ent;
1385 * Walker: Replaces Sels of frame type and
1386 * value param type entities by FrameAddress.
1387 * Links all used entities.
1389 static void lower_frame_sels_walker(ir_node *irn, void *data)
1391 lower_frame_sels_env_t *ctx = (lower_frame_sels_env_t*)data;
1394 ir_node *ptr = get_Sel_ptr(irn);
1396 if (ptr == ctx->frame) {
1397 ir_entity *ent = get_Sel_entity(irn);
1398 ir_node *bl = get_nodes_block(irn);
1401 int is_value_param = 0;
1403 if (get_entity_owner(ent) == ctx->value_tp) {
1406 /* replace by its copy from the argument type */
1407 pos = get_struct_member_index(ctx->value_tp, ent);
1408 ent = get_argument_entity(ent, ctx);
1411 nw = be_new_FrameAddr(ctx->sp_class, bl, ctx->frame, ent);
1414 /* check, if it's a param Sel and if have not seen this entity before */
1415 if (is_value_param && get_entity_link(ent) == NULL) {
1421 ARR_APP1(ent_pos_pair, ctx->value_param_list, pair);
1423 set_entity_link(ent, ctx->value_param_list);
1430 * Check if a value parameter is transmitted as a register.
1431 * This might happen if the address of an parameter is taken which is
1432 * transmitted in registers.
1434 * Note that on some architectures this case must be handled specially
1435 * because the place of the backing store is determined by their ABI.
1437 * In the default case we move the entity to the frame type and create
1438 * a backing store into the first block.
1440 static void fix_address_of_parameter_access(be_abi_irg_t *env, ir_graph *irg,
1441 ent_pos_pair *value_param_list)
1443 be_abi_call_t *call = env->call;
1444 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1445 ent_pos_pair *entry, *new_list;
1447 int i, n = ARR_LEN(value_param_list);
1450 for (i = 0; i < n; ++i) {
1451 int pos = value_param_list[i].pos;
1452 be_abi_call_arg_t *arg = get_call_arg(call, 0, pos, 1);
1455 DBG((dbg, LEVEL_2, "\targ #%d need backing store\n", pos));
1456 value_param_list[i].next = new_list;
1457 new_list = &value_param_list[i];
1460 if (new_list != NULL) {
1461 /* ok, change the graph */
1462 ir_node *start_bl = get_irg_start_block(irg);
1463 ir_node *first_bl = get_first_block_succ(start_bl);
1464 ir_node *frame, *imem, *nmem, *store, *mem, *args;
1465 optimization_state_t state;
1468 assert(first_bl && first_bl != start_bl);
1469 /* we had already removed critical edges, so the following
1470 assertion should be always true. */
1471 assert(get_Block_n_cfgpreds(first_bl) == 1);
1473 /* now create backing stores */
1474 frame = get_irg_frame(irg);
1475 imem = get_irg_initial_mem(irg);
1477 save_optimization_state(&state);
1479 nmem = new_r_Proj(get_irg_start(irg), mode_M, pn_Start_M);
1480 restore_optimization_state(&state);
1482 /* reroute all edges to the new memory source */
1483 edges_reroute(imem, nmem);
1487 args = get_irg_args(irg);
1488 for (entry = new_list; entry != NULL; entry = entry->next) {
1490 ir_type *tp = get_entity_type(entry->ent);
1491 ir_mode *mode = get_type_mode(tp);
1494 /* address for the backing store */
1495 addr = be_new_FrameAddr(arch_env->sp->reg_class, first_bl, frame, entry->ent);
1498 mem = new_r_Proj(store, mode_M, pn_Store_M);
1500 /* the backing store itself */
1501 store = new_r_Store(first_bl, mem, addr,
1502 new_r_Proj(args, mode, i), cons_none);
1504 /* the new memory Proj gets the last Proj from store */
1505 set_Proj_pred(nmem, store);
1506 set_Proj_proj(nmem, pn_Store_M);
1507 set_nodes_block(nmem, get_nodes_block(store));
1509 /* move all entities to the frame type */
1510 frame_tp = get_irg_frame_type(irg);
1511 offset = get_type_size_bytes(frame_tp);
1513 /* we will add new entities: set the layout to undefined */
1514 assert(get_type_state(frame_tp) == layout_fixed);
1515 set_type_state(frame_tp, layout_undefined);
1516 for (entry = new_list; entry != NULL; entry = entry->next) {
1517 ir_entity *ent = entry->ent;
1519 /* If the entity is still on the argument type, move it to the
1521 * This happens if the value_param type was build due to compound
1523 if (get_entity_owner(ent) != frame_tp) {
1524 ir_type *tp = get_entity_type(ent);
1525 unsigned align = get_type_alignment_bytes(tp);
1527 offset += align - 1;
1528 offset &= ~(align - 1);
1529 set_entity_owner(ent, frame_tp);
1530 /* must be automatic to set a fixed layout */
1531 set_entity_offset(ent, offset);
1532 offset += get_type_size_bytes(tp);
1535 set_type_size_bytes(frame_tp, offset);
1536 /* fix the layout again */
1537 set_type_state(frame_tp, layout_fixed);
1542 * The start block has no jump, instead it has an initial exec Proj.
1543 * The backend wants to handle all blocks the same way, so we replace
1544 * the out cfg edge with a real jump.
1546 static void fix_start_block(ir_graph *irg)
1548 ir_node *initial_X = get_irg_initial_exec(irg);
1549 ir_node *start_block = get_irg_start_block(irg);
1550 ir_node *jmp = new_r_Jmp(start_block);
1552 assert(is_Proj(initial_X));
1553 exchange(initial_X, jmp);
1554 set_irg_initial_exec(irg, new_r_Bad(irg, mode_X));
1556 /* merge start block with successor if possible */
1558 const ir_edge_t *edge;
1559 foreach_out_edge(jmp, edge) {
1560 ir_node *succ = get_edge_src_irn(edge);
1561 if (!is_Block(succ))
1564 if (get_irn_arity(succ) == 1) {
1565 exchange(succ, start_block);
1573 * Update the entity of Sels to the outer value parameters.
1575 static void update_outer_frame_sels(ir_node *irn, void *env)
1577 lower_frame_sels_env_t *ctx = (lower_frame_sels_env_t*)env;
1584 ptr = get_Sel_ptr(irn);
1585 if (! is_arg_Proj(ptr))
1587 if (get_Proj_proj(ptr) != ctx->static_link_pos)
1589 ent = get_Sel_entity(irn);
1591 if (get_entity_owner(ent) == ctx->value_tp) {
1592 /* replace by its copy from the argument type */
1593 pos = get_struct_member_index(ctx->value_tp, ent);
1594 ent = get_argument_entity(ent, ctx);
1595 set_Sel_entity(irn, ent);
1597 /* check, if we have not seen this entity before */
1598 if (get_entity_link(ent) == NULL) {
1604 ARR_APP1(ent_pos_pair, ctx->value_param_list, pair);
1606 set_entity_link(ent, ctx->value_param_list);
1612 * Fix access to outer local variables.
1614 static void fix_outer_variable_access(be_abi_irg_t *env,
1615 lower_frame_sels_env_t *ctx)
1621 for (i = get_class_n_members(ctx->frame_tp) - 1; i >= 0; --i) {
1622 ir_entity *ent = get_class_member(ctx->frame_tp, i);
1624 if (! is_method_entity(ent))
1627 irg = get_entity_irg(ent);
1632 * FIXME: find the number of the static link parameter
1633 * for now we assume 0 here
1635 ctx->static_link_pos = 0;
1637 irg_walk_graph(irg, NULL, update_outer_frame_sels, ctx);
1642 * Modify the irg itself and the frame type.
1644 static void modify_irg(ir_graph *irg)
1646 be_abi_irg_t *env = be_get_irg_abi(irg);
1647 be_abi_call_t *call = env->call;
1648 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1649 const arch_register_t *sp = arch_env->sp;
1650 ir_type *method_type = get_entity_type(get_irg_entity(irg));
1651 be_irg_t *birg = be_birg_from_irg(irg);
1652 struct obstack *obst = be_get_be_obst(irg);
1653 be_stack_layout_t *stack_layout = be_get_irg_stack_layout(irg);
1656 ir_node *new_mem_proj;
1662 unsigned frame_size;
1665 const arch_register_t *fp_reg;
1666 ir_node *frame_pointer;
1670 const ir_edge_t *edge;
1671 ir_type *arg_type, *bet_type, *tp;
1672 lower_frame_sels_env_t ctx;
1673 ir_entity **param_map;
1675 DBG((dbg, LEVEL_1, "introducing abi on %+F\n", irg));
1677 old_mem = get_irg_initial_mem(irg);
1679 irp_reserve_resources(irp, IRP_RESOURCE_ENTITY_LINK);
1681 /* set the links of all frame entities to NULL, we use it
1682 to detect if an entity is already linked in the value_param_list */
1683 tp = get_method_value_param_type(method_type);
1686 /* clear the links of the clone type, let the
1687 original entities point to its clones */
1688 for (i = get_struct_n_members(tp) - 1; i >= 0; --i) {
1689 ir_entity *mem = get_struct_member(tp, i);
1690 set_entity_link(mem, NULL);
1694 arg_type = compute_arg_type(irg, call, method_type, tp, ¶m_map);
1696 /* Convert the Sel nodes in the irg to frame addr nodes: */
1697 ctx.value_param_list = NEW_ARR_F(ent_pos_pair, 0);
1698 ctx.frame = get_irg_frame(irg);
1699 ctx.sp_class = arch_env->sp->reg_class;
1700 ctx.link_class = arch_env->link_class;
1701 ctx.frame_tp = get_irg_frame_type(irg);
1703 /* layout the stackframe now */
1704 if (get_type_state(ctx.frame_tp) == layout_undefined) {
1705 default_layout_compound_type(ctx.frame_tp);
1708 /* we will possible add new entities to the frame: set the layout to undefined */
1709 assert(get_type_state(ctx.frame_tp) == layout_fixed);
1710 set_type_state(ctx.frame_tp, layout_undefined);
1712 irg_walk_graph(irg, lower_frame_sels_walker, NULL, &ctx);
1714 /* fix the frame type layout again */
1715 set_type_state(ctx.frame_tp, layout_fixed);
1716 /* align stackframe to 4 byte */
1717 frame_size = get_type_size_bytes(ctx.frame_tp);
1718 if (frame_size % 4 != 0) {
1719 set_type_size_bytes(ctx.frame_tp, frame_size + 4 - (frame_size % 4));
1722 env->regs = pmap_create();
1724 n_params = get_method_n_params(method_type);
1725 args = OALLOCNZ(obst, ir_node*, n_params);
1728 * for inner function we must now fix access to outer frame entities.
1730 fix_outer_variable_access(env, &ctx);
1732 /* Check if a value parameter is transmitted as a register.
1733 * This might happen if the address of an parameter is taken which is
1734 * transmitted in registers.
1736 * Note that on some architectures this case must be handled specially
1737 * because the place of the backing store is determined by their ABI.
1739 * In the default case we move the entity to the frame type and create
1740 * a backing store into the first block.
1742 fix_address_of_parameter_access(env, irg, ctx.value_param_list);
1744 DEL_ARR_F(ctx.value_param_list);
1745 irp_free_resources(irp, IRP_RESOURCE_ENTITY_LINK);
1747 /* Fill the argument vector */
1748 arg_tuple = get_irg_args(irg);
1749 foreach_out_edge(arg_tuple, edge) {
1750 ir_node *irn = get_edge_src_irn(edge);
1751 if (! is_Anchor(irn)) {
1752 int nr = get_Proj_proj(irn);
1754 DBG((dbg, LEVEL_2, "\treading arg: %d -> %+F\n", nr, irn));
1758 stack_layout->sp_relative = call->flags.bits.try_omit_fp;
1759 bet_type = call->cb->get_between_type(irg);
1760 stack_frame_init(stack_layout, arg_type, bet_type,
1761 get_irg_frame_type(irg), param_map);
1763 /* Count the register params and add them to the number of Projs for the RegParams node */
1764 for (i = 0; i < n_params; ++i) {
1765 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1766 if (arg->in_reg && args[i]) {
1767 assert(arg->reg != sp && "cannot use stack pointer as parameter register");
1768 assert(i == get_Proj_proj(args[i]));
1770 /* For now, associate the register with the old Proj from Start representing that argument. */
1771 pmap_insert(env->regs, (void *) arg->reg, args[i]);
1772 DBG((dbg, LEVEL_2, "\targ #%d -> reg %s\n", i, arg->reg->name));
1776 /* Collect all callee-save registers */
1777 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
1778 const arch_register_class_t *cls = &arch_env->register_classes[i];
1779 for (j = 0; j < cls->n_regs; ++j) {
1780 const arch_register_t *reg = &cls->regs[j];
1781 if ((reg->type & arch_register_type_state) || arch_register_is_callee_save(arch_env, reg)) {
1782 pmap_insert(env->regs, (void *) reg, NULL);
1787 fp_reg = call->flags.bits.try_omit_fp ? arch_env->sp : arch_env->bp;
1788 rbitset_clear(birg->allocatable_regs, fp_reg->global_index);
1790 /* handle start block here (place a jump in the block) */
1791 fix_start_block(irg);
1793 pmap_insert(env->regs, (void *) sp, NULL);
1794 pmap_insert(env->regs, (void *) arch_env->bp, NULL);
1795 start_bl = get_irg_start_block(irg);
1796 env->start = be_new_Start(NULL, start_bl, pmap_count(env->regs) + 1);
1797 set_irg_start(irg, env->start);
1800 * make proj nodes for the callee save registers.
1801 * memorize them, since Return nodes get those as inputs.
1803 * Note, that if a register corresponds to an argument, the regs map
1804 * contains the old Proj from start for that argument.
1806 rm = ALLOCAN(reg_node_map_t, pmap_count(env->regs));
1807 reg_map_to_arr(rm, env->regs);
1808 for (i = 0, n = pmap_count(env->regs); i < n; ++i) {
1809 const arch_register_t *reg = rm[i].reg;
1810 ir_mode *mode = reg->reg_class->mode;
1812 arch_register_req_type_t add_type = arch_register_req_type_none;
1816 add_type |= arch_register_req_type_produces_sp;
1817 if (!rbitset_is_set(birg->allocatable_regs, reg->global_index)) {
1818 add_type |= arch_register_req_type_ignore;
1822 proj = new_r_Proj(env->start, mode, nr + 1);
1823 pmap_insert(env->regs, (void *) reg, proj);
1824 be_set_constr_single_reg_out(env->start, nr + 1, reg, add_type);
1825 arch_set_irn_register(proj, reg);
1827 DBG((dbg, LEVEL_2, "\tregister save proj #%d -> reg %s\n", nr, reg->name));
1830 /* create a new initial memory proj */
1831 assert(is_Proj(old_mem));
1832 arch_set_out_register_req(env->start, 0, arch_no_register_req);
1833 new_mem_proj = new_r_Proj(env->start, mode_M, 0);
1835 set_irg_initial_mem(irg, mem);
1837 env->init_sp = be_abi_reg_map_get(env->regs, sp);
1839 /* set new frame_pointer */
1840 frame_pointer = be_abi_reg_map_get(env->regs, fp_reg);
1841 set_irg_frame(irg, frame_pointer);
1843 /* rewire old mem users to new mem */
1844 exchange(old_mem, mem);
1846 /* keep the mem (for functions with an endless loop = no return) */
1849 set_irg_initial_mem(irg, mem);
1851 /* Now, introduce stack param nodes for all parameters passed on the stack */
1852 for (i = 0; i < n_params; ++i) {
1853 ir_node *arg_proj = args[i];
1854 ir_node *repl = NULL;
1856 if (arg_proj != NULL) {
1857 be_abi_call_arg_t *arg;
1858 ir_type *param_type;
1859 int nr = get_Proj_proj(arg_proj);
1862 nr = MIN(nr, n_params);
1863 arg = get_call_arg(call, 0, nr, 1);
1864 param_type = get_method_param_type(method_type, nr);
1867 repl = (ir_node*)pmap_get(env->regs, arg->reg);
1868 } else if (arg->on_stack) {
1869 ir_node *addr = be_new_FrameAddr(sp->reg_class, start_bl, frame_pointer, arg->stack_ent);
1871 /* For atomic parameters which are actually used, we create a Load node. */
1872 if (is_atomic_type(param_type) && get_irn_n_edges(args[i]) > 0) {
1873 ir_mode *mode = get_type_mode(param_type);
1874 ir_mode *load_mode = arg->load_mode;
1875 ir_node *nomem = get_irg_no_mem(irg);
1877 ir_node *load = new_r_Load(start_bl, nomem, addr, load_mode, cons_floats);
1878 repl = new_r_Proj(load, load_mode, pn_Load_res);
1880 if (mode != load_mode) {
1881 repl = new_r_Conv(start_bl, repl, mode);
1884 /* The stack parameter is not primitive (it is a struct or array),
1885 * we thus will create a node representing the parameter's address
1891 assert(repl != NULL);
1893 /* Beware: the mode of the register parameters is always the mode of the register class
1894 which may be wrong. Add Conv's then. */
1895 mode = get_irn_mode(args[i]);
1896 if (mode != get_irn_mode(repl)) {
1897 repl = new_r_Conv(get_nodes_block(repl), repl, mode);
1899 exchange(args[i], repl);
1903 /* the arg proj is not needed anymore now and should be only used by the anchor */
1904 assert(get_irn_n_edges(arg_tuple) == 1);
1905 kill_node(arg_tuple);
1906 set_irg_args(irg, new_r_Bad(irg, mode_T));
1908 /* All Return nodes hang on the End node, so look for them there. */
1909 end = get_irg_end_block(irg);
1910 for (i = 0, n = get_Block_n_cfgpreds(end); i < n; ++i) {
1911 ir_node *irn = get_Block_cfgpred(end, i);
1913 if (is_Return(irn)) {
1914 ir_node *blk = get_nodes_block(irn);
1915 ir_node *mem = get_Return_mem(irn);
1916 ir_node *ret = create_be_return(env, irn, blk, mem, get_Return_n_ress(irn));
1921 /* if we have endless loops here, n might be <= 0. Do NOT create a be_Return then,
1922 the code is dead and will never be executed. */
1925 /** Fix the state inputs of calls that still hang on unknowns */
1926 static void fix_call_state_inputs(ir_graph *irg)
1928 be_abi_irg_t *env = be_get_irg_abi(irg);
1929 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1931 arch_register_t **stateregs = NEW_ARR_F(arch_register_t*, 0);
1933 /* Collect caller save registers */
1934 n = arch_env->n_register_classes;
1935 for (i = 0; i < n; ++i) {
1937 const arch_register_class_t *cls = &arch_env->register_classes[i];
1938 for (j = 0; j < cls->n_regs; ++j) {
1939 const arch_register_t *reg = arch_register_for_index(cls, j);
1940 if (reg->type & arch_register_type_state) {
1941 ARR_APP1(arch_register_t*, stateregs, (arch_register_t *)reg);
1946 n = ARR_LEN(env->calls);
1947 n_states = ARR_LEN(stateregs);
1948 for (i = 0; i < n; ++i) {
1950 ir_node *call = env->calls[i];
1952 arity = get_irn_arity(call);
1954 /* the state reg inputs are the last n inputs of the calls */
1955 for (s = 0; s < n_states; ++s) {
1956 int inp = arity - n_states + s;
1957 const arch_register_t *reg = stateregs[s];
1958 ir_node *regnode = be_abi_reg_map_get(env->regs, reg);
1960 set_irn_n(call, inp, regnode);
1964 DEL_ARR_F(stateregs);
1968 * Create a trampoline entity for the given method.
1970 static ir_entity *create_trampoline(be_main_env_t *be, ir_entity *method)
1972 ir_type *type = get_entity_type(method);
1973 ident *old_id = get_entity_ld_ident(method);
1974 ident *id = id_mangle3("", old_id, "$stub");
1975 ir_type *parent = be->pic_trampolines_type;
1976 ir_entity *ent = new_entity(parent, old_id, type);
1977 set_entity_ld_ident(ent, id);
1978 set_entity_visibility(ent, ir_visibility_private);
1984 * Returns the trampoline entity for the given method.
1986 static ir_entity *get_trampoline(be_main_env_t *env, ir_entity *method)
1988 ir_entity *result = (ir_entity*)pmap_get(env->ent_trampoline_map, method);
1989 if (result == NULL) {
1990 result = create_trampoline(env, method);
1991 pmap_insert(env->ent_trampoline_map, method, result);
1997 static ir_entity *create_pic_symbol(be_main_env_t *be, ir_entity *entity)
1999 ident *old_id = get_entity_ld_ident(entity);
2000 ident *id = id_mangle3("", old_id, "$non_lazy_ptr");
2001 ir_type *e_type = get_entity_type(entity);
2002 ir_type *type = new_type_pointer(e_type);
2003 ir_type *parent = be->pic_symbols_type;
2004 ir_entity *ent = new_entity(parent, old_id, type);
2005 set_entity_ld_ident(ent, id);
2006 set_entity_visibility(ent, ir_visibility_private);
2011 static ir_entity *get_pic_symbol(be_main_env_t *env, ir_entity *entity)
2013 ir_entity *result = (ir_entity*)pmap_get(env->ent_pic_symbol_map, entity);
2014 if (result == NULL) {
2015 result = create_pic_symbol(env, entity);
2016 pmap_insert(env->ent_pic_symbol_map, entity, result);
2025 * Returns non-zero if a given entity can be accessed using a relative address.
2027 static int can_address_relative(ir_entity *entity)
2029 return get_entity_visibility(entity) != ir_visibility_external
2030 && !(get_entity_linkage(entity) & IR_LINKAGE_MERGE);
2033 static ir_node *get_pic_base(ir_graph *irg)
2035 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
2036 if (arch_env->impl->get_pic_base == NULL)
2038 return arch_env->impl->get_pic_base(irg);
2041 /** patches SymConsts to work in position independent code */
2042 static void fix_pic_symconsts(ir_node *node, void *data)
2044 ir_graph *irg = get_irn_irg(node);
2045 be_main_env_t *be = be_get_irg_main_env(irg);
2055 arity = get_irn_arity(node);
2056 for (i = 0; i < arity; ++i) {
2058 ir_node *pred = get_irn_n(node, i);
2060 ir_entity *pic_symbol;
2061 ir_node *pic_symconst;
2063 if (!is_SymConst(pred))
2066 entity = get_SymConst_entity(pred);
2067 block = get_nodes_block(pred);
2069 /* calls can jump to relative addresses, so we can directly jump to
2070 the (relatively) known call address or the trampoline */
2071 if (i == 1 && is_Call(node)) {
2072 ir_entity *trampoline;
2073 ir_node *trampoline_const;
2075 if (can_address_relative(entity))
2078 dbgi = get_irn_dbg_info(pred);
2079 trampoline = get_trampoline(be, entity);
2080 trampoline_const = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
2082 set_irn_n(node, i, trampoline_const);
2086 /* everything else is accessed relative to EIP */
2087 mode = get_irn_mode(pred);
2088 pic_base = get_pic_base(irg);
2090 /* all ok now for locally constructed stuff */
2091 if (can_address_relative(entity)) {
2092 ir_node *add = new_r_Add(block, pic_base, pred, mode);
2094 /* make sure the walker doesn't visit this add again */
2095 mark_irn_visited(add);
2096 set_irn_n(node, i, add);
2100 /* get entry from pic symbol segment */
2101 dbgi = get_irn_dbg_info(pred);
2102 pic_symbol = get_pic_symbol(be, entity);
2103 pic_symconst = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
2105 add = new_r_Add(block, pic_base, pic_symconst, mode);
2106 mark_irn_visited(add);
2108 /* we need an extra indirection for global data outside our current
2109 module. The loads are always safe and can therefore float
2110 and need no memory input */
2111 load = new_r_Load(block, get_irg_no_mem(irg), add, mode, cons_floats);
2112 load_res = new_r_Proj(load, mode, pn_Load_res);
2114 set_irn_n(node, i, load_res);
2118 void be_abi_introduce(ir_graph *irg)
2120 be_abi_irg_t *env = XMALLOCZ(be_abi_irg_t);
2121 ir_node *old_frame = get_irg_frame(irg);
2122 be_options_t *options = be_get_irg_options(irg);
2123 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
2124 ir_entity *entity = get_irg_entity(irg);
2125 ir_type *method_type = get_entity_type(entity);
2126 be_irg_t *birg = be_birg_from_irg(irg);
2127 struct obstack *obst = &birg->obst;
2128 ir_node *dummy = new_r_Dummy(irg,
2129 arch_env->sp->reg_class->mode);
2132 /* determine allocatable registers */
2133 assert(birg->allocatable_regs == NULL);
2134 birg->allocatable_regs = rbitset_obstack_alloc(obst, arch_env->n_registers);
2135 for (r = 0; r < arch_env->n_registers; ++r) {
2136 const arch_register_t *reg = &arch_env->registers[r];
2137 if ( !(reg->type & arch_register_type_ignore)) {
2138 rbitset_set(birg->allocatable_regs, r);
2142 /* break here if backend provides a custom API.
2143 * Note: we shouldn't have to setup any be_abi_irg_t* stuff at all,
2144 * but need more cleanup to make this work
2146 be_set_irg_abi(irg, env);
2148 be_omit_fp = options->omit_fp;
2150 env->keep_map = pmap_create();
2151 env->call = be_abi_call_new(arch_env->sp->reg_class);
2152 arch_env_get_call_abi(arch_env, method_type, env->call);
2154 env->init_sp = dummy;
2155 env->calls = NEW_ARR_F(ir_node*, 0);
2160 irg_walk_graph(irg, fix_pic_symconsts, NULL, env);
2163 /* Lower all call nodes in the IRG. */
2166 /* Process the IRG */
2169 /* fix call inputs for state registers */
2170 fix_call_state_inputs(irg);
2172 /* We don't need the keep map anymore. */
2173 pmap_destroy(env->keep_map);
2174 env->keep_map = NULL;
2176 /* calls array is not needed anymore */
2177 DEL_ARR_F(env->calls);
2180 /* reroute the stack origin of the calls to the true stack origin. */
2181 exchange(dummy, env->init_sp);
2182 exchange(old_frame, get_irg_frame(irg));
2184 pmap_destroy(env->regs);
2188 void be_abi_free(ir_graph *irg)
2190 be_abi_irg_t *env = be_get_irg_abi(irg);
2192 if (env->call != NULL)
2193 be_abi_call_free(env->call);
2194 assert(env->regs == NULL);
2197 be_set_irg_abi(irg, NULL);
2200 void be_put_allocatable_regs(const ir_graph *irg,
2201 const arch_register_class_t *cls, bitset_t *bs)
2203 be_irg_t *birg = be_birg_from_irg(irg);
2204 unsigned *allocatable_regs = birg->allocatable_regs;
2207 assert(bitset_size(bs) == cls->n_regs);
2208 bitset_clear_all(bs);
2209 for (i = 0; i < cls->n_regs; ++i) {
2210 const arch_register_t *reg = &cls->regs[i];
2211 if (rbitset_is_set(allocatable_regs, reg->global_index))
2216 unsigned be_get_n_allocatable_regs(const ir_graph *irg,
2217 const arch_register_class_t *cls)
2219 bitset_t *bs = bitset_alloca(cls->n_regs);
2220 be_put_allocatable_regs(irg, cls, bs);
2221 return bitset_popcount(bs);
2224 void be_set_allocatable_regs(const ir_graph *irg,
2225 const arch_register_class_t *cls,
2226 unsigned *raw_bitset)
2228 be_irg_t *birg = be_birg_from_irg(irg);
2229 unsigned *allocatable_regs = birg->allocatable_regs;
2232 rbitset_clear_all(raw_bitset, cls->n_regs);
2233 for (i = 0; i < cls->n_regs; ++i) {
2234 const arch_register_t *reg = &cls->regs[i];
2235 if (rbitset_is_set(allocatable_regs, reg->global_index))
2236 rbitset_set(raw_bitset, i);
2240 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_abi)
2241 void be_init_abi(void)
2243 FIRM_DBG_REGISTER(dbg, "firm.be.abi");