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"
46 #include "raw_bitset.h"
57 #include "bessaconstr.h"
59 #include "betranshlp.h"
61 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
63 typedef struct be_abi_call_arg_t {
64 unsigned is_res : 1; /**< 1: the call argument is a return value. 0: it's a call parameter. */
65 unsigned in_reg : 1; /**< 1: this argument is transmitted in registers. */
66 unsigned on_stack : 1; /**< 1: this argument is transmitted on the stack. */
67 unsigned callee : 1; /**< 1: someone called us. 0: We call another function */
70 const arch_register_t *reg;
73 unsigned alignment; /**< stack alignment */
74 unsigned space_before; /**< allocate space before */
75 unsigned space_after; /**< allocate space after */
78 struct be_abi_call_t {
79 be_abi_call_flags_t flags; /**< Flags describing the ABI behavior on calls */
80 int pop; /**< number of bytes the stack frame is shrinked by the callee on return. */
81 const be_abi_callbacks_t *cb;
82 ir_type *between_type;
84 const arch_register_class_t *cls_addr; /**< register class of the call address */
88 * The ABI information for the current graph.
91 be_abi_call_t *call; /**< The ABI call information. */
93 ir_node *init_sp; /**< The node representing the stack pointer
94 at the start of the function. */
96 ir_node *start; /**< The be_Start params node. */
97 pmap *regs; /**< A map of all callee-save and ignore regs to
98 their Projs to the RegParams node. */
99 int start_block_bias; /**< The stack bias at the end of the start block. */
101 pmap *keep_map; /**< mapping blocks to keep nodes. */
103 ir_node **calls; /**< flexible array containing all be_Call nodes */
106 static ir_heights_t *ir_heights;
108 /** Flag: if set, try to omit the frame pointer in all routines. */
109 static int be_omit_fp = 1;
111 static ir_node *be_abi_reg_map_get(pmap *map, const arch_register_t *reg)
113 return (ir_node*)pmap_get(map, reg);
116 static void be_abi_reg_map_set(pmap *map, const arch_register_t* reg,
119 pmap_insert(map, reg, node);
123 * Check if the given register is callee save, ie. will be saved by the callee.
125 static bool arch_register_is_callee_save(
126 const arch_env_t *arch_env,
127 const arch_register_t *reg)
129 if (arch_env->impl->register_saved_by)
130 return arch_env->impl->register_saved_by(reg, /*callee=*/1);
135 * Check if the given register is caller save, ie. must be saved by the caller.
137 static bool arch_register_is_caller_save(
138 const arch_env_t *arch_env,
139 const arch_register_t *reg)
141 if (arch_env->impl->register_saved_by)
142 return arch_env->impl->register_saved_by(reg, /*callee=*/0);
149 _ ____ ___ ____ _ _ _ _
150 / \ | __ )_ _| / ___|__ _| | | |__ __ _ ___| | _____
151 / _ \ | _ \| | | | / _` | | | '_ \ / _` |/ __| |/ / __|
152 / ___ \| |_) | | | |__| (_| | | | |_) | (_| | (__| <\__ \
153 /_/ \_\____/___| \____\__,_|_|_|_.__/ \__,_|\___|_|\_\___/
155 These callbacks are used by the backend to set the parameters
156 for a specific call type.
160 * Set compare function: compares two ABI call object arguments.
162 static int cmp_call_arg(const void *a, const void *b, size_t n)
164 const be_abi_call_arg_t *p = (const be_abi_call_arg_t*)a;
165 const be_abi_call_arg_t *q = (const be_abi_call_arg_t*)b;
167 return !(p->is_res == q->is_res && p->pos == q->pos && p->callee == q->callee);
171 * Get an ABI call object argument.
173 * @param call the abi call
174 * @param is_res true for call results, false for call arguments
175 * @param pos position of the argument
176 * @param callee context type - if we are callee or caller
178 static be_abi_call_arg_t *get_call_arg(be_abi_call_t *call, int is_res, int pos, int callee)
180 be_abi_call_arg_t arg;
183 memset(&arg, 0, sizeof(arg));
188 hash = is_res * 128 + pos;
190 return (be_abi_call_arg_t*)set_find(call->params, &arg, sizeof(arg), hash);
194 * Set an ABI call object argument.
196 static void remember_call_arg(be_abi_call_arg_t *arg, be_abi_call_t *call, be_abi_context_t context)
198 unsigned hash = arg->is_res * 128 + arg->pos;
199 if (context & ABI_CONTEXT_CALLEE) {
201 set_insert(call->params, arg, sizeof(*arg), hash);
203 if (context & ABI_CONTEXT_CALLER) {
205 set_insert(call->params, arg, sizeof(*arg), hash);
209 /* Set the flags for a call. */
210 void be_abi_call_set_flags(be_abi_call_t *call, be_abi_call_flags_t flags, const be_abi_callbacks_t *cb)
216 /* Sets the number of bytes the stackframe is shrinked by the callee on return */
217 void be_abi_call_set_pop(be_abi_call_t *call, int pop)
223 /* Set register class for call address */
224 void be_abi_call_set_call_address_reg_class(be_abi_call_t *call, const arch_register_class_t *cls)
226 call->cls_addr = cls;
230 void be_abi_call_param_stack(be_abi_call_t *call, int arg_pos,
231 ir_mode *load_mode, unsigned alignment,
232 unsigned space_before, unsigned space_after,
233 be_abi_context_t context)
235 be_abi_call_arg_t arg;
236 memset(&arg, 0, sizeof(arg));
237 assert(alignment > 0 && "Alignment must be greater than 0");
239 arg.load_mode = load_mode;
240 arg.alignment = alignment;
241 arg.space_before = space_before;
242 arg.space_after = space_after;
246 remember_call_arg(&arg, call, context);
249 void be_abi_call_param_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
251 be_abi_call_arg_t arg;
252 memset(&arg, 0, sizeof(arg));
259 remember_call_arg(&arg, call, context);
262 void be_abi_call_res_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
264 be_abi_call_arg_t arg;
265 memset(&arg, 0, sizeof(arg));
272 remember_call_arg(&arg, call, context);
275 /* Get the flags of a ABI call object. */
276 be_abi_call_flags_t be_abi_call_get_flags(const be_abi_call_t *call)
282 * Constructor for a new ABI call object.
284 * @param cls_addr register class of the call address
286 * @return the new ABI call object
288 static be_abi_call_t *be_abi_call_new(const arch_register_class_t *cls_addr)
290 be_abi_call_t *call = XMALLOCZ(be_abi_call_t);
293 call->params = new_set(cmp_call_arg, 16);
295 call->cls_addr = cls_addr;
297 call->flags.bits.try_omit_fp = be_omit_fp;
303 * Destructor for an ABI call object.
305 static void be_abi_call_free(be_abi_call_t *call)
307 del_set(call->params);
312 * Initializes the frame layout from parts
314 * @param frame the stack layout that will be initialized
315 * @param args the stack argument layout type
316 * @param between the between layout type
317 * @param locals the method frame type
318 * @param param_map an array mapping method argument positions to the stack argument type
320 * @return the initialized stack layout
322 static be_stack_layout_t *stack_frame_init(be_stack_layout_t *frame, ir_type *args,
323 ir_type *between, ir_type *locals,
324 ir_entity *param_map[])
326 frame->arg_type = args;
327 frame->between_type = between;
328 frame->frame_type = locals;
329 frame->initial_offset = 0;
330 frame->initial_bias = 0;
331 frame->order[1] = between;
332 frame->param_map = param_map;
334 /* typical decreasing stack: locals have the
335 * lowest addresses, arguments the highest */
336 frame->order[0] = locals;
337 frame->order[2] = args;
348 Adjustment of the calls inside a graph.
353 * Transform a call node into a be_Call node.
355 * @param env The ABI environment for the current irg.
356 * @param irn The call node.
357 * @param curr_sp The stack pointer node to use.
358 * @return The stack pointer after the call.
360 static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
362 ir_graph *irg = get_irn_irg(irn);
363 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
364 ir_type *call_tp = get_Call_type(irn);
365 ir_node *call_ptr = get_Call_ptr(irn);
366 size_t n_params = get_method_n_params(call_tp);
367 ir_node *curr_mem = get_Call_mem(irn);
368 ir_node *bl = get_nodes_block(irn);
370 const arch_register_t *sp = arch_env->sp;
371 be_abi_call_t *call = be_abi_call_new(sp->reg_class);
372 ir_mode *mach_mode = sp->reg_class->mode;
373 int no_alloc = call->flags.bits.frame_is_setup_on_call;
374 int n_res = get_method_n_ress(call_tp);
375 int do_seq = call->flags.bits.store_args_sequential && !no_alloc;
377 ir_node *res_proj = NULL;
378 int n_reg_params = 0;
379 int n_stack_params = 0;
382 const arch_register_t **states = NEW_ARR_F(const arch_register_t*, 0);
383 const arch_register_t **destroyed_regs = NEW_ARR_F(const arch_register_t*, 0);
387 int n_reg_results = 0;
388 const ir_edge_t *edge;
390 int *stack_param_idx;
392 int throws_exception;
397 /* Let the isa fill out the abi description for that call node. */
398 arch_env_get_call_abi(arch_env, call_tp, call);
400 /* Insert code to put the stack arguments on the stack. */
401 assert(get_Call_n_params(irn) == n_params);
402 stack_param_idx = ALLOCAN(int, n_params);
403 for (p = 0; p < n_params; ++p) {
404 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
407 int arg_size = get_type_size_bytes(get_method_param_type(call_tp, p));
409 stack_size += round_up2(arg->space_before, arg->alignment);
410 stack_size += round_up2(arg_size, arg->alignment);
411 stack_size += round_up2(arg->space_after, arg->alignment);
413 stack_param_idx[n_stack_params++] = p;
417 /* Collect all arguments which are passed in registers. */
418 reg_param_idxs = ALLOCAN(int, n_params);
419 for (p = 0; p < n_params; ++p) {
420 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
421 if (arg && arg->in_reg) {
422 reg_param_idxs[n_reg_params++] = p;
427 * If the stack is decreasing and we do not want to store sequentially,
428 * or someone else allocated the call frame
429 * we allocate as much space on the stack all parameters need, by
430 * moving the stack pointer along the stack's direction.
432 * Note: we also have to do this for stack_size == 0, because we may have
433 * to adjust stack alignment for the call.
435 if (!do_seq && !no_alloc) {
436 curr_sp = be_new_IncSP(sp, bl, curr_sp, stack_size, 1);
439 dbgi = get_irn_dbg_info(irn);
440 /* If there are some parameters which shall be passed on the stack. */
441 if (n_stack_params > 0) {
443 ir_node **in = ALLOCAN(ir_node*, n_stack_params+1);
446 /* push params in reverse direction because stack grows downwards */
448 for (i = 0; i < n_stack_params >> 1; ++i) {
449 int other = n_stack_params - i - 1;
450 int tmp = stack_param_idx[i];
451 stack_param_idx[i] = stack_param_idx[other];
452 stack_param_idx[other] = tmp;
456 curr_mem = get_Call_mem(irn);
458 in[n_in++] = curr_mem;
461 for (i = 0; i < n_stack_params; ++i) {
462 int p = stack_param_idx[i];
463 be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
464 ir_node *param = get_Call_param(irn, p);
465 ir_node *addr = curr_sp;
467 ir_type *param_type = get_method_param_type(call_tp, p);
468 int param_size = get_type_size_bytes(param_type) + arg->space_after;
471 * If we wanted to build the arguments sequentially,
472 * the stack pointer for the next must be incremented,
473 * and the memory value propagated.
477 addr = curr_sp = be_new_IncSP(sp, bl, curr_sp,
478 param_size + arg->space_before, 0);
479 add_irn_dep(curr_sp, curr_mem);
481 curr_ofs += arg->space_before;
482 curr_ofs = round_up2(curr_ofs, arg->alignment);
484 /* Make the expression to compute the argument's offset. */
486 ir_mode *constmode = mach_mode;
487 if (mode_is_reference(mach_mode)) {
490 addr = new_r_Const_long(irg, constmode, curr_ofs);
491 addr = new_r_Add(bl, curr_sp, addr, mach_mode);
495 /* Insert a store for primitive arguments. */
496 if (is_atomic_type(param_type)) {
497 ir_node *nomem = get_irg_no_mem(irg);
498 ir_node *mem_input = do_seq ? curr_mem : nomem;
499 ir_node *store = new_rd_Store(dbgi, bl, mem_input, addr, param, cons_none);
500 mem = new_r_Proj(store, mode_M, pn_Store_M);
502 /* Make a mem copy for compound arguments. */
505 assert(mode_is_reference(get_irn_mode(param)));
506 copy = new_rd_CopyB(dbgi, bl, curr_mem, addr, param, param_type);
507 mem = new_r_Proj(copy, mode_M, pn_CopyB_M);
510 curr_ofs += param_size;
518 /* We need the sync only, if we didn't build the stores sequentially. */
520 if (n_stack_params >= 1) {
521 curr_mem = new_r_Sync(bl, n_in, in);
523 curr_mem = get_Call_mem(irn);
528 /* Put caller save into the destroyed set and state registers in the states
530 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
532 const arch_register_class_t *cls = &arch_env->register_classes[i];
533 for (j = 0; j < cls->n_regs; ++j) {
534 const arch_register_t *reg = arch_register_for_index(cls, j);
536 /* even if destroyed all is specified, neither SP nor FP are
537 * destroyed (else bad things will happen) */
538 if (reg == arch_env->sp || reg == arch_env->bp)
541 if (reg->type & arch_register_type_state) {
542 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
543 ARR_APP1(const arch_register_t*, states, reg);
544 /* we're already in the destroyed set so no need for further
548 if (arch_register_is_caller_save(arch_env, reg)) {
549 if (!(reg->type & arch_register_type_ignore)) {
550 ARR_APP1(const arch_register_t*, destroyed_regs, reg);
556 /* search the largest result proj number */
557 res_projs = ALLOCANZ(ir_node*, n_res);
559 foreach_out_edge(irn, edge) {
560 const ir_edge_t *res_edge;
561 ir_node *irn = get_edge_src_irn(edge);
563 if (!is_Proj(irn) || get_Proj_proj(irn) != pn_Call_T_result)
566 foreach_out_edge(irn, res_edge) {
568 ir_node *res = get_edge_src_irn(res_edge);
570 assert(is_Proj(res));
572 proj = get_Proj_proj(res);
573 assert(proj < n_res);
574 assert(res_projs[proj] == NULL);
575 res_projs[proj] = res;
581 /** TODO: this is not correct for cases where return values are passed
582 * on the stack, but no known ABI does this currently...
584 n_reg_results = n_res;
587 in = ALLOCAN(ir_node*, n_reg_params + ARR_LEN(states));
589 /* make the back end call node and set its register requirements. */
590 for (i = 0; i < n_reg_params; ++i) {
591 in[n_ins++] = get_Call_param(irn, reg_param_idxs[i]);
594 /* add state registers ins */
595 for (s = 0; s < ARR_LEN(states); ++s) {
596 const arch_register_t *reg = states[s];
597 const arch_register_class_t *cls = reg->reg_class;
598 ir_node *regnode = new_r_Unknown(irg, cls->mode);
599 in[n_ins++] = regnode;
601 assert(n_ins == (int) (n_reg_params + ARR_LEN(states)));
603 /* ins collected, build the call */
604 throws_exception = ir_throws_exception(irn);
605 if (env->call->flags.bits.call_has_imm && is_SymConst(call_ptr)) {
607 low_call = be_new_Call(dbgi, irg, bl, curr_mem, curr_sp, curr_sp,
608 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
609 n_ins, in, get_Call_type(irn));
610 be_Call_set_entity(low_call, get_SymConst_entity(call_ptr));
613 low_call = be_new_Call(dbgi, irg, bl, curr_mem, curr_sp, call_ptr,
614 n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
615 n_ins, in, get_Call_type(irn));
617 ir_set_throws_exception(low_call, throws_exception);
618 be_Call_set_pop(low_call, call->pop);
620 /* put the call into the list of all calls for later processing */
621 ARR_APP1(ir_node *, env->calls, low_call);
623 /* create new stack pointer */
624 curr_sp = new_r_Proj(low_call, get_irn_mode(curr_sp), pn_be_Call_sp);
625 be_set_constr_single_reg_out(low_call, pn_be_Call_sp, sp,
626 arch_register_req_type_ignore | arch_register_req_type_produces_sp);
627 arch_set_irn_register(curr_sp, sp);
629 /* now handle results */
630 for (i = 0; i < n_res; ++i) {
631 ir_node *proj = res_projs[i];
632 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
633 long pn = i + pn_be_Call_first_res;
635 /* returns values on stack not supported yet */
639 shift the proj number to the right, since we will drop the
640 unspeakable Proj_T from the Call. Therefore, all real argument
641 Proj numbers must be increased by pn_be_Call_first_res
643 pn = i + pn_be_Call_first_res;
646 ir_type *res_type = get_method_res_type(call_tp, i);
647 ir_mode *mode = get_type_mode(res_type);
648 proj = new_r_Proj(low_call, mode, pn);
651 set_Proj_pred(proj, low_call);
652 set_Proj_proj(proj, pn);
656 /* remove register from destroyed regs */
658 size_t n = ARR_LEN(destroyed_regs);
659 for (j = 0; j < n; ++j) {
660 if (destroyed_regs[j] == arg->reg) {
661 destroyed_regs[j] = destroyed_regs[n-1];
662 ARR_SHRINKLEN(destroyed_regs,n-1);
670 Set the register class of the call address to
671 the backend provided class (default: stack pointer class)
673 be_node_set_reg_class_in(low_call, n_be_Call_ptr, call->cls_addr);
675 DBG((dbg, LEVEL_3, "\tcreated backend call %+F\n", low_call));
677 /* Set the register classes and constraints of the Call parameters. */
678 for (i = 0; i < n_reg_params; ++i) {
679 int index = reg_param_idxs[i];
680 be_abi_call_arg_t *arg = get_call_arg(call, 0, index, 0);
681 assert(arg->reg != NULL);
683 be_set_constr_single_reg_in(low_call, n_be_Call_first_arg + i,
684 arg->reg, arch_register_req_type_none);
687 /* Set the register constraints of the results. */
688 for (i = 0; i < n_res; ++i) {
689 ir_node *proj = res_projs[i];
690 const be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 0);
691 int pn = get_Proj_proj(proj);
694 be_set_constr_single_reg_out(low_call, pn, arg->reg,
695 arch_register_req_type_none);
696 arch_set_irn_register(proj, arg->reg);
698 exchange(irn, low_call);
700 /* kill the ProjT node */
701 if (res_proj != NULL) {
705 /* Make additional projs for the caller save registers
706 and the Keep node which keeps them alive. */
712 int curr_res_proj = pn_be_Call_first_res + n_reg_results;
715 n_ins = ARR_LEN(destroyed_regs) + n_reg_results + 1;
716 in = ALLOCAN(ir_node *, n_ins);
718 /* also keep the stack pointer */
719 set_irn_link(curr_sp, (void*) sp);
722 for (d = 0; d < ARR_LEN(destroyed_regs); ++d) {
723 const arch_register_t *reg = destroyed_regs[d];
724 ir_node *proj = new_r_Proj(low_call, reg->reg_class->mode, curr_res_proj);
726 /* memorize the register in the link field. we need afterwards to set the register class of the keep correctly. */
727 be_set_constr_single_reg_out(low_call, curr_res_proj, reg,
728 arch_register_req_type_none);
729 arch_set_irn_register(proj, reg);
731 set_irn_link(proj, (void*) reg);
736 for (i = 0; i < n_reg_results; ++i) {
737 ir_node *proj = res_projs[i];
738 const arch_register_t *reg = arch_get_irn_register(proj);
739 set_irn_link(proj, (void*) reg);
744 /* create the Keep for the caller save registers */
745 keep = be_new_Keep(bl, n, in);
746 for (i = 0; i < n; ++i) {
747 const arch_register_t *reg = (const arch_register_t*)get_irn_link(in[i]);
748 be_node_set_reg_class_in(keep, i, arch_register_get_class(reg));
752 /* Clean up the stack. */
753 assert(stack_size >= call->pop);
754 stack_size -= call->pop;
756 if (stack_size > 0) {
757 ir_node *mem_proj = NULL;
759 foreach_out_edge(low_call, edge) {
760 ir_node *irn = get_edge_src_irn(edge);
761 if (is_Proj(irn) && get_Proj_proj(irn) == pn_Call_M) {
768 mem_proj = new_r_Proj(low_call, mode_M, pn_be_Call_M);
769 keep_alive(mem_proj);
772 /* Clean up the stack frame or revert alignment fixes if we allocated it */
774 curr_sp = be_new_IncSP(sp, bl, curr_sp, -stack_size, 0);
777 be_abi_call_free(call);
780 DEL_ARR_F(destroyed_regs);
786 * Adjust the size of a node representing a stack alloc or free for the minimum stack alignment.
788 * @param alignment the minimum stack alignment
789 * @param size the node containing the non-aligned size
790 * @param block the block where new nodes are allocated on
791 * @param dbg debug info for new nodes
793 * @return a node representing the aligned size
795 static ir_node *adjust_alloc_size(unsigned stack_alignment, ir_node *size,
796 ir_node *block, dbg_info *dbg)
798 if (stack_alignment > 1) {
804 assert(is_po2(stack_alignment));
806 mode = get_irn_mode(size);
807 tv = new_tarval_from_long(stack_alignment-1, mode);
808 irg = get_Block_irg(block);
809 mask = new_r_Const(irg, tv);
810 size = new_rd_Add(dbg, block, size, mask, mode);
812 tv = new_tarval_from_long(-(long)stack_alignment, mode);
813 mask = new_r_Const(irg, tv);
814 size = new_rd_And(dbg, block, size, mask, mode);
820 * The alloca is transformed into a back end alloca node and connected to the stack nodes.
822 static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp)
824 ir_node *block = get_nodes_block(alloc);
825 ir_graph *irg = get_Block_irg(block);
826 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
827 ir_node *alloc_mem = NULL;
828 ir_node *alloc_res = NULL;
829 ir_type *type = get_Alloc_type(alloc);
832 const ir_edge_t *edge;
837 unsigned stack_alignment;
839 /* all non-stack Alloc nodes should already be lowered before the backend */
840 assert(get_Alloc_where(alloc) == stack_alloc);
842 foreach_out_edge(alloc, edge) {
843 ir_node *irn = get_edge_src_irn(edge);
845 assert(is_Proj(irn));
846 switch (get_Proj_proj(irn)) {
858 /* Beware: currently Alloc nodes without a result might happen,
859 only escape analysis kills them and this phase runs only for object
860 oriented source. We kill the Alloc here. */
861 if (alloc_res == NULL && alloc_mem) {
862 exchange(alloc_mem, get_Alloc_mem(alloc));
866 dbg = get_irn_dbg_info(alloc);
867 count = get_Alloc_count(alloc);
869 /* we might need to multiply the count with the element size */
870 if (type != firm_unknown_type && get_type_size_bytes(type) != 1) {
871 ir_mode *mode = get_irn_mode(count);
872 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type),
874 ir_node *cnst = new_rd_Const(dbg, irg, tv);
875 size = new_rd_Mul(dbg, block, count, cnst, mode);
880 /* The stack pointer will be modified in an unknown manner.
881 We cannot omit it. */
882 env->call->flags.bits.try_omit_fp = 0;
884 stack_alignment = 1 << arch_env->stack_alignment;
885 size = adjust_alloc_size(stack_alignment, size, block, dbg);
886 new_alloc = be_new_AddSP(arch_env->sp, block, curr_sp, size);
887 set_irn_dbg_info(new_alloc, dbg);
889 if (alloc_mem != NULL) {
893 addsp_mem = new_r_Proj(new_alloc, mode_M, pn_be_AddSP_M);
895 /* We need to sync the output mem of the AddSP with the input mem
896 edge into the alloc node. */
897 ins[0] = get_Alloc_mem(alloc);
899 sync = new_r_Sync(block, 2, ins);
901 exchange(alloc_mem, sync);
904 exchange(alloc, new_alloc);
906 /* fix projnum of alloca res */
907 set_Proj_proj(alloc_res, pn_be_AddSP_res);
909 curr_sp = new_r_Proj(new_alloc, get_irn_mode(curr_sp), pn_be_AddSP_sp);
916 * The Free is transformed into a back end free node and connected to the stack nodes.
918 static ir_node *adjust_free(be_abi_irg_t *env, ir_node *free, ir_node *curr_sp)
920 ir_node *block = get_nodes_block(free);
921 ir_graph *irg = get_irn_irg(free);
922 ir_type *type = get_Free_type(free);
923 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
924 ir_mode *sp_mode = arch_env->sp->reg_class->mode;
925 dbg_info *dbg = get_irn_dbg_info(free);
926 ir_node *subsp, *mem, *res, *size, *sync;
928 unsigned stack_alignment;
930 /* all non-stack-alloc Free nodes should already be lowered before the
932 assert(get_Free_where(free) == stack_alloc);
934 /* we might need to multiply the size with the element size */
935 if (type != firm_unknown_type && get_type_size_bytes(type) != 1) {
936 ir_tarval *tv = new_tarval_from_long(get_type_size_bytes(type), mode_Iu);
937 ir_node *cnst = new_rd_Const(dbg, irg, tv);
938 ir_node *mul = new_rd_Mul(dbg, block, get_Free_count(free),
942 size = get_Free_count(free);
945 stack_alignment = 1 << arch_env->stack_alignment;
946 size = adjust_alloc_size(stack_alignment, size, block, dbg);
948 /* The stack pointer will be modified in an unknown manner.
949 We cannot omit it. */
950 env->call->flags.bits.try_omit_fp = 0;
951 subsp = be_new_SubSP(arch_env->sp, block, curr_sp, size);
952 set_irn_dbg_info(subsp, dbg);
954 mem = new_r_Proj(subsp, mode_M, pn_be_SubSP_M);
955 res = new_r_Proj(subsp, sp_mode, pn_be_SubSP_sp);
957 /* we need to sync the memory */
958 in[0] = get_Free_mem(free);
960 sync = new_r_Sync(block, 2, in);
962 /* and make the AddSP dependent on the former memory */
963 add_irn_dep(subsp, get_Free_mem(free));
966 exchange(free, sync);
973 * Check if a node is somehow data dependent on another one.
974 * both nodes must be in the same basic block.
975 * @param n1 The first node.
976 * @param n2 The second node.
977 * @return 1, if n1 is data dependent (transitively) on n2, 0 if not.
979 static int dependent_on(ir_node *n1, ir_node *n2)
981 assert(get_nodes_block(n1) == get_nodes_block(n2));
983 return heights_reachable_in_block(ir_heights, n1, n2);
986 static int cmp_call_dependency(const void *c1, const void *c2)
988 ir_node *n1 = *(ir_node **) c1;
989 ir_node *n2 = *(ir_node **) c2;
993 Classical qsort() comparison function behavior:
994 0 if both elements are equal
995 1 if second is "smaller" that first
996 -1 if first is "smaller" that second
998 if (dependent_on(n1, n2))
1001 if (dependent_on(n2, n1))
1004 /* The nodes have no depth order, but we need a total order because qsort()
1007 * Additionally, we need to respect transitive dependencies. Consider a
1008 * Call a depending on Call b and an independent Call c.
1009 * We MUST NOT order c > a and b > c. */
1010 h1 = get_irn_height(ir_heights, n1);
1011 h2 = get_irn_height(ir_heights, n2);
1012 if (h1 < h2) return -1;
1013 if (h1 > h2) return 1;
1014 /* Same height, so use a random (but stable) order */
1015 return get_irn_idx(n1) - get_irn_idx(n2);
1019 * Walker: links all Call/Alloc/Free nodes to the Block they are contained.
1020 * Clears the irg_is_leaf flag if a Call is detected.
1022 static void link_ops_in_block_walker(ir_node *irn, void *data)
1024 be_abi_irg_t *env = (be_abi_irg_t*)data;
1025 unsigned code = get_irn_opcode(irn);
1027 if (code == iro_Call ||
1028 (code == iro_Alloc && get_Alloc_where(irn) == stack_alloc) ||
1029 (code == iro_Free && get_Free_where(irn) == stack_alloc)) {
1030 ir_node *bl = get_nodes_block(irn);
1031 void *save = get_irn_link(bl);
1033 if (code == iro_Call)
1034 env->call->flags.bits.irg_is_leaf = 0;
1036 set_irn_link(irn, save);
1037 set_irn_link(bl, irn);
1040 if (code == iro_Builtin && get_Builtin_kind(irn) == ir_bk_return_address) {
1041 ir_node *param = get_Builtin_param(irn, 0);
1042 ir_tarval *tv = get_Const_tarval(param);
1043 unsigned long value = get_tarval_long(tv);
1044 /* use ebp, so the climbframe algo works... */
1046 env->call->flags.bits.try_omit_fp = 0;
1053 * Process all Call/Alloc/Free nodes inside a basic block.
1054 * Note that the link field of the block must contain a linked list of all
1055 * nodes inside the Block. We first order this list according to data dependency
1056 * and that connect the nodes together.
1058 static void process_ops_in_block(ir_node *bl, void *data)
1060 be_abi_irg_t *env = (be_abi_irg_t*)data;
1061 ir_node *curr_sp = env->init_sp;
1068 for (irn = (ir_node*)get_irn_link(bl); irn != NULL;
1069 irn = (ir_node*)get_irn_link(irn)) {
1073 nodes = ALLOCAN(ir_node*, n_nodes);
1074 for (irn = (ir_node*)get_irn_link(bl), n = 0; irn != NULL;
1075 irn = (ir_node*)get_irn_link(irn), ++n) {
1079 /* If there were call nodes in the block. */
1084 /* order the call nodes according to data dependency */
1085 qsort(nodes, n_nodes, sizeof(nodes[0]), cmp_call_dependency);
1087 for (i = n_nodes - 1; i >= 0; --i) {
1088 ir_node *irn = nodes[i];
1090 DBG((dbg, LEVEL_3, "\tprocessing call %+F\n", irn));
1091 switch (get_irn_opcode(irn)) {
1094 /* The stack pointer will be modified due to a call. */
1095 env->call->flags.bits.try_omit_fp = 0;
1097 curr_sp = adjust_call(env, irn, curr_sp);
1100 if (get_Alloc_where(irn) == stack_alloc)
1101 curr_sp = adjust_alloc(env, irn, curr_sp);
1104 if (get_Free_where(irn) == stack_alloc)
1105 curr_sp = adjust_free(env, irn, curr_sp);
1108 panic("invalid call");
1112 /* Keep the last stack state in the block by tying it to Keep node,
1113 * the proj from calls is already kept */
1114 if (curr_sp != env->init_sp &&
1115 !(is_Proj(curr_sp) && be_is_Call(get_Proj_pred(curr_sp)))) {
1117 keep = be_new_Keep(bl, 1, nodes);
1118 pmap_insert(env->keep_map, bl, keep);
1122 set_irn_link(bl, curr_sp);
1126 * Adjust all call nodes in the graph to the ABI conventions.
1128 static void process_calls(ir_graph *irg)
1130 be_abi_irg_t *abi = be_get_irg_abi(irg);
1132 abi->call->flags.bits.irg_is_leaf = 1;
1133 irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, abi);
1135 ir_heights = heights_new(irg);
1136 irg_block_walk_graph(irg, NULL, process_ops_in_block, abi);
1137 heights_free(ir_heights);
1141 * Computes the stack argument layout type.
1142 * Changes a possibly allocated value param type by moving
1143 * entities to the stack layout type.
1145 * @param call the current call ABI
1146 * @param method_type the method type
1147 * @param param_map an array mapping method arguments to the stack layout
1150 * @return the stack argument layout type
1152 static ir_type *compute_arg_type(ir_graph *irg, be_abi_call_t *call,
1153 ir_type *method_type, ir_entity ***param_map)
1155 struct obstack *obst = be_get_be_obst(irg);
1156 ir_type *frame_type = get_irg_frame_type(irg);
1157 size_t n_params = get_method_n_params(method_type);
1158 size_t n_frame_members = get_compound_n_members(frame_type);
1159 ir_entity *va_start_entity = NULL;
1167 *param_map = map = OALLOCNZ(obst, ir_entity*, n_params);
1168 res = new_type_struct(new_id_from_chars("arg_type", 8));
1170 /* collect existing entities for value_param_types */
1171 for (f = n_frame_members; f > 0; ) {
1172 ir_entity *entity = get_compound_member(frame_type, --f);
1175 set_entity_link(entity, NULL);
1176 if (!is_parameter_entity(entity))
1178 num = get_entity_parameter_number(entity);
1179 if (num == IR_VA_START_PARAMETER_NUMBER) {
1180 /* move entity to new arg_type */
1181 set_entity_owner(entity, res);
1182 va_start_entity = entity;
1185 assert(num < n_params);
1186 if (map[num] != NULL)
1187 panic("multiple entities for parameter %u in %+F found", f, irg);
1189 if (num != n_params && !get_call_arg(call, 0, num, 1)->on_stack) {
1190 /* don't move this entity */
1195 /* move entity to new arg_type */
1196 set_entity_owner(entity, res);
1199 for (i = 0; i < n_params; ++i) {
1200 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1201 ir_type *param_type = get_method_param_type(method_type, i);
1204 if (!arg->on_stack) {
1208 if (entity == NULL) {
1209 /* create a new entity */
1210 entity = new_parameter_entity(res, i, param_type);
1212 ofs += arg->space_before;
1213 ofs = round_up2(ofs, arg->alignment);
1214 set_entity_offset(entity, ofs);
1215 ofs += arg->space_after;
1216 ofs += get_type_size_bytes(param_type);
1217 arg->stack_ent = entity;
1219 if (va_start_entity != NULL) {
1220 set_entity_offset(va_start_entity, ofs);
1222 set_type_size_bytes(res, ofs);
1223 set_type_state(res, layout_fixed);
1229 const arch_register_t *reg;
1233 static int cmp_regs(const void *a, const void *b)
1235 const reg_node_map_t *p = (const reg_node_map_t*)a;
1236 const reg_node_map_t *q = (const reg_node_map_t*)b;
1238 if (p->reg->reg_class == q->reg->reg_class)
1239 return p->reg->index - q->reg->index;
1241 return p->reg->reg_class < q->reg->reg_class ? -1 : +1;
1244 static void reg_map_to_arr(reg_node_map_t *res, pmap *reg_map)
1247 size_t n = pmap_count(reg_map);
1250 foreach_pmap(reg_map, ent) {
1251 res[i].reg = (const arch_register_t*)ent->key;
1252 res[i].irn = (ir_node*)ent->value;
1256 qsort(res, n, sizeof(res[0]), cmp_regs);
1260 * Creates a be_Return for a Return node.
1262 * @param @env the abi environment
1263 * @param irn the Return node or NULL if there was none
1264 * @param bl the block where the be_Retun should be placed
1265 * @param mem the current memory
1266 * @param n_res number of return results
1268 static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl,
1269 ir_node *mem, int n_res)
1271 be_abi_call_t *call = env->call;
1272 ir_graph *irg = get_Block_irg(bl);
1273 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1275 pmap *reg_map = pmap_create();
1276 ir_node *keep = (ir_node*)pmap_get(env->keep_map, bl);
1283 const arch_register_t **regs;
1287 get the valid stack node in this block.
1288 If we had a call in that block there is a Keep constructed by process_calls()
1289 which points to the last stack modification in that block. we'll use
1290 it then. Else we use the stack from the start block and let
1291 the ssa construction fix the usage.
1293 stack = be_abi_reg_map_get(env->regs, arch_env->sp);
1295 stack = get_irn_n(keep, 0);
1297 remove_End_keepalive(get_irg_end(irg), keep);
1300 /* Insert results for Return into the register map. */
1301 for (i = 0; i < n_res; ++i) {
1302 ir_node *res = get_Return_res(irn, i);
1303 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1304 assert(arg->in_reg && "return value must be passed in register");
1305 pmap_insert(reg_map, (void *) arg->reg, res);
1308 /* Add uses of the callee save registers. */
1309 foreach_pmap(env->regs, ent) {
1310 const arch_register_t *reg = (const arch_register_t*)ent->key;
1311 if ((reg->type & arch_register_type_ignore) || arch_register_is_callee_save(arch_env, reg))
1312 pmap_insert(reg_map, ent->key, ent->value);
1315 be_abi_reg_map_set(reg_map, arch_env->sp, stack);
1318 Maximum size of the in array for Return nodes is
1319 return args + callee save/ignore registers + memory + stack pointer
1321 in_max = pmap_count(reg_map) + n_res + 2;
1323 in = ALLOCAN(ir_node*, in_max);
1324 regs = ALLOCAN(arch_register_t const*, in_max);
1327 in[1] = be_abi_reg_map_get(reg_map, arch_env->sp);
1329 regs[1] = arch_env->sp;
1332 /* clear SP entry, since it has already been grown. */
1333 pmap_insert(reg_map, (void *) arch_env->sp, NULL);
1334 for (i = 0; i < n_res; ++i) {
1335 be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);
1337 in[n] = be_abi_reg_map_get(reg_map, arg->reg);
1338 regs[n++] = arg->reg;
1340 /* Clear the map entry to mark the register as processed. */
1341 be_abi_reg_map_set(reg_map, arg->reg, NULL);
1344 /* grow the rest of the stuff. */
1345 foreach_pmap(reg_map, ent) {
1347 in[n] = (ir_node*)ent->value;
1348 regs[n++] = (const arch_register_t*)ent->key;
1352 /* The in array for the new back end return is now ready. */
1354 dbgi = get_irn_dbg_info(irn);
1358 /* we have to pop the shadow parameter in in case of struct returns */
1360 ret = be_new_Return(dbgi, irg, bl, n_res, pop, n, in);
1362 /* Set the register classes of the return's parameter accordingly. */
1363 for (i = 0; i < n; ++i) {
1364 if (regs[i] == NULL)
1367 be_set_constr_single_reg_in(ret, i, regs[i], arch_register_req_type_none);
1370 /* Free the space of the Epilog's in array and the register <-> proj map. */
1371 pmap_destroy(reg_map);
1376 typedef struct lower_frame_sels_env_t {
1377 ir_node *frame; /**< the current frame */
1378 const arch_register_class_t *sp_class; /**< register class of the stack pointer */
1379 const arch_register_class_t *link_class; /**< register class of the link pointer */
1380 ir_type *frame_tp; /**< the frame type */
1381 int static_link_pos; /**< argument number of the hidden static link */
1382 } lower_frame_sels_env_t;
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 nw = be_new_FrameAddr(ctx->sp_class, bl, ctx->frame, ent);
1408 * The start block has no jump, instead it has an initial exec Proj.
1409 * The backend wants to handle all blocks the same way, so we replace
1410 * the out cfg edge with a real jump.
1412 static void fix_start_block(ir_graph *irg)
1414 ir_node *initial_X = get_irg_initial_exec(irg);
1415 ir_node *start_block = get_irg_start_block(irg);
1416 ir_node *jmp = new_r_Jmp(start_block);
1418 assert(is_Proj(initial_X));
1419 exchange(initial_X, jmp);
1420 set_irg_initial_exec(irg, new_r_Bad(irg, mode_X));
1422 /* merge start block with successor if possible */
1424 const ir_edge_t *edge;
1425 foreach_out_edge(jmp, edge) {
1426 ir_node *succ = get_edge_src_irn(edge);
1427 if (!is_Block(succ))
1430 if (get_irn_arity(succ) == 1) {
1431 exchange(succ, start_block);
1439 * Modify the irg itself and the frame type.
1441 static void modify_irg(ir_graph *irg)
1443 be_abi_irg_t *env = be_get_irg_abi(irg);
1444 be_abi_call_t *call = env->call;
1445 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1446 const arch_register_t *sp = arch_env->sp;
1447 ir_type *method_type = get_entity_type(get_irg_entity(irg));
1448 be_irg_t *birg = be_birg_from_irg(irg);
1449 struct obstack *obst = be_get_be_obst(irg);
1450 be_stack_layout_t *stack_layout = be_get_irg_stack_layout(irg);
1453 ir_node *new_mem_proj;
1459 unsigned frame_size;
1462 const arch_register_t *fp_reg;
1463 ir_node *frame_pointer;
1467 const ir_edge_t *edge;
1468 ir_type *arg_type, *bet_type;
1469 lower_frame_sels_env_t ctx;
1470 ir_entity **param_map;
1472 DBG((dbg, LEVEL_1, "introducing abi on %+F\n", irg));
1474 old_mem = get_irg_initial_mem(irg);
1476 irp_reserve_resources(irp, IRP_RESOURCE_ENTITY_LINK);
1478 arg_type = compute_arg_type(irg, call, method_type, ¶m_map);
1480 /* Convert the Sel nodes in the irg to frame addr nodes: */
1481 ctx.frame = get_irg_frame(irg);
1482 ctx.sp_class = arch_env->sp->reg_class;
1483 ctx.link_class = arch_env->link_class;
1484 ctx.frame_tp = get_irg_frame_type(irg);
1486 /* layout the stackframe now */
1487 if (get_type_state(ctx.frame_tp) == layout_undefined) {
1488 default_layout_compound_type(ctx.frame_tp);
1491 /* align stackframe to 4 byte */
1492 frame_size = get_type_size_bytes(ctx.frame_tp);
1493 if (frame_size % 4 != 0) {
1494 set_type_size_bytes(ctx.frame_tp, frame_size + 4 - (frame_size % 4));
1497 env->regs = pmap_create();
1499 n_params = get_method_n_params(method_type);
1500 args = OALLOCNZ(obst, ir_node*, n_params);
1502 be_add_parameter_entity_stores(irg);
1504 irg_walk_graph(irg, lower_frame_sels_walker, NULL, &ctx);
1506 irp_free_resources(irp, IRP_RESOURCE_ENTITY_LINK);
1508 /* Fill the argument vector */
1509 arg_tuple = get_irg_args(irg);
1510 foreach_out_edge(arg_tuple, edge) {
1511 ir_node *irn = get_edge_src_irn(edge);
1512 if (! is_Anchor(irn)) {
1513 int nr = get_Proj_proj(irn);
1515 DBG((dbg, LEVEL_2, "\treading arg: %d -> %+F\n", nr, irn));
1519 stack_layout->sp_relative = call->flags.bits.try_omit_fp;
1520 bet_type = call->cb->get_between_type(irg);
1521 stack_frame_init(stack_layout, arg_type, bet_type,
1522 get_irg_frame_type(irg), param_map);
1524 /* Count the register params and add them to the number of Projs for the RegParams node */
1525 for (i = 0; i < n_params; ++i) {
1526 be_abi_call_arg_t *arg = get_call_arg(call, 0, i, 1);
1527 if (arg->in_reg && args[i]) {
1528 assert(arg->reg != sp && "cannot use stack pointer as parameter register");
1529 assert(i == get_Proj_proj(args[i]));
1531 /* For now, associate the register with the old Proj from Start representing that argument. */
1532 pmap_insert(env->regs, (void *) arg->reg, args[i]);
1533 DBG((dbg, LEVEL_2, "\targ #%d -> reg %s\n", i, arg->reg->name));
1537 /* Collect all callee-save registers */
1538 for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
1539 const arch_register_class_t *cls = &arch_env->register_classes[i];
1540 for (j = 0; j < cls->n_regs; ++j) {
1541 const arch_register_t *reg = &cls->regs[j];
1542 if ((reg->type & arch_register_type_state) || arch_register_is_callee_save(arch_env, reg)) {
1543 pmap_insert(env->regs, (void *) reg, NULL);
1548 fp_reg = call->flags.bits.try_omit_fp ? arch_env->sp : arch_env->bp;
1549 rbitset_clear(birg->allocatable_regs, fp_reg->global_index);
1551 /* handle start block here (place a jump in the block) */
1552 fix_start_block(irg);
1554 pmap_insert(env->regs, (void *) sp, NULL);
1555 pmap_insert(env->regs, (void *) arch_env->bp, NULL);
1556 start_bl = get_irg_start_block(irg);
1557 env->start = be_new_Start(NULL, start_bl, pmap_count(env->regs) + 1);
1558 set_irg_start(irg, env->start);
1561 * make proj nodes for the callee save registers.
1562 * memorize them, since Return nodes get those as inputs.
1564 * Note, that if a register corresponds to an argument, the regs map
1565 * contains the old Proj from start for that argument.
1567 rm = ALLOCAN(reg_node_map_t, pmap_count(env->regs));
1568 reg_map_to_arr(rm, env->regs);
1569 for (i = 0, n = pmap_count(env->regs); i < n; ++i) {
1570 const arch_register_t *reg = rm[i].reg;
1571 ir_mode *mode = reg->reg_class->mode;
1573 arch_register_req_type_t add_type = arch_register_req_type_none;
1577 add_type |= arch_register_req_type_produces_sp;
1578 if (!rbitset_is_set(birg->allocatable_regs, reg->global_index)) {
1579 add_type |= arch_register_req_type_ignore;
1583 proj = new_r_Proj(env->start, mode, nr + 1);
1584 pmap_insert(env->regs, (void *) reg, proj);
1585 be_set_constr_single_reg_out(env->start, nr + 1, reg, add_type);
1586 arch_set_irn_register(proj, reg);
1588 DBG((dbg, LEVEL_2, "\tregister save proj #%d -> reg %s\n", nr, reg->name));
1591 /* create a new initial memory proj */
1592 assert(is_Proj(old_mem));
1593 arch_set_irn_register_req_out(env->start, 0, arch_no_register_req);
1594 new_mem_proj = new_r_Proj(env->start, mode_M, 0);
1596 set_irg_initial_mem(irg, mem);
1598 env->init_sp = be_abi_reg_map_get(env->regs, sp);
1600 /* set new frame_pointer */
1601 frame_pointer = be_abi_reg_map_get(env->regs, fp_reg);
1602 set_irg_frame(irg, frame_pointer);
1604 /* rewire old mem users to new mem */
1605 exchange(old_mem, mem);
1607 /* keep the mem (for functions with an endless loop = no return) */
1610 set_irg_initial_mem(irg, mem);
1612 /* Now, introduce stack param nodes for all parameters passed on the stack */
1613 for (i = 0; i < n_params; ++i) {
1614 ir_node *arg_proj = args[i];
1615 ir_node *repl = NULL;
1617 if (arg_proj != NULL) {
1618 be_abi_call_arg_t *arg;
1619 ir_type *param_type;
1620 int nr = get_Proj_proj(arg_proj);
1623 nr = MIN(nr, n_params);
1624 arg = get_call_arg(call, 0, nr, 1);
1625 param_type = get_method_param_type(method_type, nr);
1628 repl = (ir_node*)pmap_get(env->regs, arg->reg);
1629 } else if (arg->on_stack) {
1630 ir_node *addr = be_new_FrameAddr(sp->reg_class, start_bl, frame_pointer, arg->stack_ent);
1632 /* For atomic parameters which are actually used, we create a Load node. */
1633 if (is_atomic_type(param_type) && get_irn_n_edges(args[i]) > 0) {
1634 ir_mode *mode = get_type_mode(param_type);
1635 ir_mode *load_mode = arg->load_mode;
1636 ir_node *nomem = get_irg_no_mem(irg);
1638 ir_node *load = new_r_Load(start_bl, nomem, addr, load_mode, cons_floats);
1639 repl = new_r_Proj(load, load_mode, pn_Load_res);
1641 if (mode != load_mode) {
1642 repl = new_r_Conv(start_bl, repl, mode);
1645 /* The stack parameter is not primitive (it is a struct or array),
1646 * we thus will create a node representing the parameter's address
1652 assert(repl != NULL);
1654 /* Beware: the mode of the register parameters is always the mode of the register class
1655 which may be wrong. Add Conv's then. */
1656 mode = get_irn_mode(args[i]);
1657 if (mode != get_irn_mode(repl)) {
1658 repl = new_r_Conv(get_nodes_block(repl), repl, mode);
1660 exchange(args[i], repl);
1664 /* the arg proj is not needed anymore now and should be only used by the anchor */
1665 assert(get_irn_n_edges(arg_tuple) == 1);
1666 kill_node(arg_tuple);
1667 set_irg_args(irg, new_r_Bad(irg, mode_T));
1669 /* All Return nodes hang on the End node, so look for them there. */
1670 end = get_irg_end_block(irg);
1671 for (i = 0, n = get_Block_n_cfgpreds(end); i < n; ++i) {
1672 ir_node *irn = get_Block_cfgpred(end, i);
1674 if (is_Return(irn)) {
1675 ir_node *blk = get_nodes_block(irn);
1676 ir_node *mem = get_Return_mem(irn);
1677 ir_node *ret = create_be_return(env, irn, blk, mem, get_Return_n_ress(irn));
1682 /* if we have endless loops here, n might be <= 0. Do NOT create a be_Return then,
1683 the code is dead and will never be executed. */
1686 /** Fix the state inputs of calls that still hang on unknowns */
1687 static void fix_call_state_inputs(ir_graph *irg)
1689 be_abi_irg_t *env = be_get_irg_abi(irg);
1690 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1692 arch_register_t **stateregs = NEW_ARR_F(arch_register_t*, 0);
1694 /* Collect caller save registers */
1695 n = arch_env->n_register_classes;
1696 for (i = 0; i < n; ++i) {
1698 const arch_register_class_t *cls = &arch_env->register_classes[i];
1699 for (j = 0; j < cls->n_regs; ++j) {
1700 const arch_register_t *reg = arch_register_for_index(cls, j);
1701 if (reg->type & arch_register_type_state) {
1702 ARR_APP1(arch_register_t*, stateregs, (arch_register_t *)reg);
1707 n = ARR_LEN(env->calls);
1708 n_states = ARR_LEN(stateregs);
1709 for (i = 0; i < n; ++i) {
1711 ir_node *call = env->calls[i];
1713 arity = get_irn_arity(call);
1715 /* the state reg inputs are the last n inputs of the calls */
1716 for (s = 0; s < n_states; ++s) {
1717 int inp = arity - n_states + s;
1718 const arch_register_t *reg = stateregs[s];
1719 ir_node *regnode = be_abi_reg_map_get(env->regs, reg);
1721 set_irn_n(call, inp, regnode);
1725 DEL_ARR_F(stateregs);
1729 * Create a trampoline entity for the given method.
1731 static ir_entity *create_trampoline(be_main_env_t *be, ir_entity *method)
1733 ir_type *type = get_entity_type(method);
1734 ident *old_id = get_entity_ld_ident(method);
1735 ident *id = id_mangle3("", old_id, "$stub");
1736 ir_type *parent = be->pic_trampolines_type;
1737 ir_entity *ent = new_entity(parent, old_id, type);
1738 set_entity_ld_ident(ent, id);
1739 set_entity_visibility(ent, ir_visibility_private);
1745 * Returns the trampoline entity for the given method.
1747 static ir_entity *get_trampoline(be_main_env_t *env, ir_entity *method)
1749 ir_entity *result = (ir_entity*)pmap_get(env->ent_trampoline_map, method);
1750 if (result == NULL) {
1751 result = create_trampoline(env, method);
1752 pmap_insert(env->ent_trampoline_map, method, result);
1758 static ir_entity *create_pic_symbol(be_main_env_t *be, ir_entity *entity)
1760 ident *old_id = get_entity_ld_ident(entity);
1761 ident *id = id_mangle3("", old_id, "$non_lazy_ptr");
1762 ir_type *e_type = get_entity_type(entity);
1763 ir_type *type = new_type_pointer(e_type);
1764 ir_type *parent = be->pic_symbols_type;
1765 ir_entity *ent = new_entity(parent, old_id, type);
1766 set_entity_ld_ident(ent, id);
1767 set_entity_visibility(ent, ir_visibility_private);
1772 static ir_entity *get_pic_symbol(be_main_env_t *env, ir_entity *entity)
1774 ir_entity *result = (ir_entity*)pmap_get(env->ent_pic_symbol_map, entity);
1775 if (result == NULL) {
1776 result = create_pic_symbol(env, entity);
1777 pmap_insert(env->ent_pic_symbol_map, entity, result);
1786 * Returns non-zero if a given entity can be accessed using a relative address.
1788 static int can_address_relative(ir_entity *entity)
1790 return get_entity_visibility(entity) != ir_visibility_external
1791 && !(get_entity_linkage(entity) & IR_LINKAGE_MERGE);
1794 static ir_node *get_pic_base(ir_graph *irg)
1796 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1797 if (arch_env->impl->get_pic_base == NULL)
1799 return arch_env->impl->get_pic_base(irg);
1802 /** patches SymConsts to work in position independent code */
1803 static void fix_pic_symconsts(ir_node *node, void *data)
1805 ir_graph *irg = get_irn_irg(node);
1806 be_main_env_t *be = be_get_irg_main_env(irg);
1816 arity = get_irn_arity(node);
1817 for (i = 0; i < arity; ++i) {
1819 ir_node *pred = get_irn_n(node, i);
1821 ir_entity *pic_symbol;
1822 ir_node *pic_symconst;
1824 if (!is_SymConst(pred))
1827 entity = get_SymConst_entity(pred);
1828 block = get_nodes_block(pred);
1830 /* calls can jump to relative addresses, so we can directly jump to
1831 the (relatively) known call address or the trampoline */
1832 if (i == 1 && is_Call(node)) {
1833 ir_entity *trampoline;
1834 ir_node *trampoline_const;
1836 if (can_address_relative(entity))
1839 dbgi = get_irn_dbg_info(pred);
1840 trampoline = get_trampoline(be, entity);
1841 trampoline_const = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
1843 set_irn_n(node, i, trampoline_const);
1847 /* everything else is accessed relative to EIP */
1848 mode = get_irn_mode(pred);
1849 pic_base = get_pic_base(irg);
1851 /* all ok now for locally constructed stuff */
1852 if (can_address_relative(entity)) {
1853 ir_node *add = new_r_Add(block, pic_base, pred, mode);
1855 /* make sure the walker doesn't visit this add again */
1856 mark_irn_visited(add);
1857 set_irn_n(node, i, add);
1861 /* get entry from pic symbol segment */
1862 dbgi = get_irn_dbg_info(pred);
1863 pic_symbol = get_pic_symbol(be, entity);
1864 pic_symconst = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
1866 add = new_r_Add(block, pic_base, pic_symconst, mode);
1867 mark_irn_visited(add);
1869 /* we need an extra indirection for global data outside our current
1870 module. The loads are always safe and can therefore float
1871 and need no memory input */
1872 load = new_r_Load(block, get_irg_no_mem(irg), add, mode, cons_floats);
1873 load_res = new_r_Proj(load, mode, pn_Load_res);
1875 set_irn_n(node, i, load_res);
1879 void be_abi_introduce(ir_graph *irg)
1881 be_abi_irg_t *env = XMALLOCZ(be_abi_irg_t);
1882 ir_node *old_frame = get_irg_frame(irg);
1883 be_options_t *options = be_get_irg_options(irg);
1884 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
1885 ir_entity *entity = get_irg_entity(irg);
1886 ir_type *method_type = get_entity_type(entity);
1887 be_irg_t *birg = be_birg_from_irg(irg);
1888 struct obstack *obst = &birg->obst;
1889 ir_node *dummy = new_r_Dummy(irg,
1890 arch_env->sp->reg_class->mode);
1893 /* determine allocatable registers */
1894 assert(birg->allocatable_regs == NULL);
1895 birg->allocatable_regs = rbitset_obstack_alloc(obst, arch_env->n_registers);
1896 for (r = 0; r < arch_env->n_registers; ++r) {
1897 const arch_register_t *reg = &arch_env->registers[r];
1898 if ( !(reg->type & arch_register_type_ignore)) {
1899 rbitset_set(birg->allocatable_regs, r);
1903 /* break here if backend provides a custom API.
1904 * Note: we shouldn't have to setup any be_abi_irg_t* stuff at all,
1905 * but need more cleanup to make this work
1907 be_set_irg_abi(irg, env);
1909 be_omit_fp = options->omit_fp;
1911 env->keep_map = pmap_create();
1912 env->call = be_abi_call_new(arch_env->sp->reg_class);
1913 arch_env_get_call_abi(arch_env, method_type, env->call);
1915 env->init_sp = dummy;
1916 env->calls = NEW_ARR_F(ir_node*, 0);
1921 irg_walk_graph(irg, fix_pic_symconsts, NULL, env);
1924 /* Lower all call nodes in the IRG. */
1927 /* Process the IRG */
1930 /* fix call inputs for state registers */
1931 fix_call_state_inputs(irg);
1933 /* We don't need the keep map anymore. */
1934 pmap_destroy(env->keep_map);
1935 env->keep_map = NULL;
1937 /* calls array is not needed anymore */
1938 DEL_ARR_F(env->calls);
1941 /* reroute the stack origin of the calls to the true stack origin. */
1942 exchange(dummy, env->init_sp);
1943 exchange(old_frame, get_irg_frame(irg));
1945 pmap_destroy(env->regs);
1949 void be_abi_free(ir_graph *irg)
1951 be_abi_irg_t *env = be_get_irg_abi(irg);
1953 if (env->call != NULL)
1954 be_abi_call_free(env->call);
1955 assert(env->regs == NULL);
1958 be_set_irg_abi(irg, NULL);
1961 void be_put_allocatable_regs(const ir_graph *irg,
1962 const arch_register_class_t *cls, bitset_t *bs)
1964 be_irg_t *birg = be_birg_from_irg(irg);
1965 unsigned *allocatable_regs = birg->allocatable_regs;
1968 assert(bitset_size(bs) == cls->n_regs);
1969 bitset_clear_all(bs);
1970 for (i = 0; i < cls->n_regs; ++i) {
1971 const arch_register_t *reg = &cls->regs[i];
1972 if (rbitset_is_set(allocatable_regs, reg->global_index))
1977 unsigned be_get_n_allocatable_regs(const ir_graph *irg,
1978 const arch_register_class_t *cls)
1980 bitset_t *bs = bitset_alloca(cls->n_regs);
1981 be_put_allocatable_regs(irg, cls, bs);
1982 return bitset_popcount(bs);
1985 void be_set_allocatable_regs(const ir_graph *irg,
1986 const arch_register_class_t *cls,
1987 unsigned *raw_bitset)
1989 be_irg_t *birg = be_birg_from_irg(irg);
1990 unsigned *allocatable_regs = birg->allocatable_regs;
1993 rbitset_clear_all(raw_bitset, cls->n_regs);
1994 for (i = 0; i < cls->n_regs; ++i) {
1995 const arch_register_t *reg = &cls->regs[i];
1996 if (rbitset_is_set(allocatable_regs, reg->global_index))
1997 rbitset_set(raw_bitset, i);
2001 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_abi)
2002 void be_init_abi(void)
2004 FIRM_DBG_REGISTER(dbg, "firm.be.abi");