2 * Copyright (C) 1995-2007 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 The main arm backend driver file.
23 * @author Oliver Richter, Tobias Gneist
30 #include <libcore/lc_opts.h>
31 #include <libcore/lc_opts_enum.h>
33 #include "pseudo_irg.h"
45 #include "../bearch_t.h" /* the general register allocator interface */
46 #include "../benode_t.h"
47 #include "../belower.h"
48 #include "../besched_t.h"
51 #include "../bemachine.h"
52 #include "../beilpsched.h"
53 #include "../bemodule.h"
54 #include "../beirg_t.h"
55 #include "../bespillslots.h"
56 #include "../begnuas.h"
58 #include "bearch_arm_t.h"
60 #include "arm_new_nodes.h" /* arm nodes interface */
61 #include "gen_arm_regalloc_if.h" /* the generated interface (register type and class defenitions) */
62 #include "arm_transform.h"
63 #include "arm_emitter.h"
64 #include "arm_map_regs.h"
66 #define DEBUG_MODULE "firm.be.arm.isa"
68 /* TODO: ugly, but we need it to get access to the registers assigned to Phi nodes */
69 static set *cur_reg_set = NULL;
71 /**************************************************
74 * _ __ ___ __ _ __ _| | | ___ ___ _| |_
75 * | '__/ _ \/ _` | / _` | | |/ _ \ / __| | | _|
76 * | | | __/ (_| | | (_| | | | (_) | (__ | | |
77 * |_| \___|\__, | \__,_|_|_|\___/ \___| |_|_|
80 **************************************************/
83 * Return register requirements for a arm node.
84 * If the node returns a tuple (mode_T) then the proj's
85 * will be asked for this information.
88 arch_register_req_t *arm_get_irn_reg_req(const void *self, const ir_node *node,
91 long node_pos = pos == -1 ? 0 : pos;
92 ir_mode *mode = get_irn_mode(node);
95 if (is_Block(node) || mode == mode_X) {
96 return arch_no_register_req;
99 if (mode == mode_T && pos < 0) {
100 return arch_no_register_req;
105 return arch_no_register_req;
108 return arch_no_register_req;
111 node_pos = (pos == -1) ? get_Proj_proj(node) : pos;
112 node = skip_Proj_const(node);
115 /* get requirements for our own nodes */
116 if (is_arm_irn(node)) {
117 const arch_register_req_t *req;
119 req = get_arm_in_req(node, pos);
121 req = get_arm_out_req(node, node_pos);
127 /* unknown should be transformed by now */
128 assert(!is_Unknown(node));
129 return arch_no_register_req;
132 static void arm_set_irn_reg(const void *self, ir_node *irn,
133 const arch_register_t *reg)
138 if (get_irn_mode(irn) == mode_X) {
143 pos = get_Proj_proj(irn);
144 irn = skip_Proj(irn);
147 if (is_arm_irn(irn)) {
148 const arch_register_t **slots;
150 slots = get_arm_slots(irn);
154 /* here we set the registers for the Phi nodes */
155 arm_set_firm_reg(irn, reg, cur_reg_set);
159 static const arch_register_t *arm_get_irn_reg(const void *self,
163 const arch_register_t *reg = NULL;
168 if (get_irn_mode(irn) == mode_X) {
172 pos = get_Proj_proj(irn);
173 irn = skip_Proj_const(irn);
176 if (is_arm_irn(irn)) {
177 const arch_register_t **slots;
178 slots = get_arm_slots(irn);
182 reg = arm_get_firm_reg(irn, cur_reg_set);
188 static arch_irn_class_t arm_classify(const void *self, const ir_node *irn)
191 irn = skip_Proj_const(irn);
194 return arch_irn_class_branch;
196 else if (is_arm_irn(irn)) {
197 return arch_irn_class_normal;
203 static arch_irn_flags_t arm_get_flags(const void *self, const ir_node *irn)
206 irn = skip_Proj_const(irn);
208 if (is_arm_irn(irn)) {
209 return get_arm_flags(irn);
211 else if (is_Unknown(irn)) {
212 return arch_irn_flags_ignore;
218 static ir_entity *arm_get_frame_entity(const void *self, const ir_node *irn)
222 /* TODO: return the entity assigned to the frame */
226 static void arm_set_frame_entity(const void *self, ir_node *irn, ir_entity *ent)
231 /* TODO: set the entity assigned to the frame */
235 * This function is called by the generic backend to correct offsets for
236 * nodes accessing the stack.
238 static void arm_set_stack_bias(const void *self, ir_node *irn, int bias)
243 /* TODO: correct offset if irn accesses the stack */
246 static int arm_get_sp_bias(const void *self, const ir_node *irn)
253 /* fill register allocator interface */
255 static const arch_irn_ops_if_t arm_irn_ops_if = {
261 arm_get_frame_entity,
262 arm_set_frame_entity,
265 NULL, /* get_inverse */
266 NULL, /* get_op_estimated_cost */
267 NULL, /* possible_memory_operand */
268 NULL, /* perform_memory_operand */
271 arm_irn_ops_t arm_irn_ops = {
278 /**************************************************
281 * ___ ___ __| | ___ __ _ ___ _ __ _| |_
282 * / __/ _ \ / _` |/ _ \/ _` |/ _ \ '_ \ | | _|
283 * | (_| (_) | (_| | __/ (_| | __/ | | | | | |
284 * \___\___/ \__,_|\___|\__, |\___|_| |_| |_|_|
287 **************************************************/
290 * Transforms the standard Firm graph into
293 static void arm_prepare_graph(void *self) {
294 arm_code_gen_t *cg = self;
296 /* transform nodes into assembler instructions */
297 arm_transform_graph(cg);
299 /* do local optimizations (mainly CSE) */
300 local_optimize_graph(cg->irg);
303 be_dump(cg->irg, "-transformed", dump_ir_block_graph_sched);
305 /* do code placement, to optimize the position of constants */
309 be_dump(cg->irg, "-place", dump_ir_block_graph_sched);
313 * Called immediately before emit phase.
315 static void arm_finish_irg(void *self)
318 /* TODO: - fix offsets for nodes accessing stack
325 * These are some hooks which must be filled but are probably not needed.
327 static void arm_before_sched(void *self)
330 /* Some stuff you need to do after scheduling but before register allocation */
333 static void arm_before_ra(void *self)
336 /* Some stuff you need to do immediately after register allocation */
340 * We transform Spill and Reload here. This needs to be done before
341 * stack biasing otherwise we would miss the corrected offset for these nodes.
343 static void arm_after_ra(void *self)
345 arm_code_gen_t *cg = self;
346 be_coalesce_spillslots(cg->birg);
350 * Emits the code, closes the output file and frees
351 * the code generator interface.
353 static void arm_emit_and_done(void *self) {
354 arm_code_gen_t *cg = self;
355 ir_graph *irg = cg->irg;
357 arm_gen_routine(cg, irg);
361 /* de-allocate code generator */
362 del_set(cg->reg_set);
367 * Move a double floating point value into an integer register.
368 * Place the move operation into block bl.
370 * Handle some special cases here:
371 * 1.) A constant: simply split into two
372 * 2.) A load: simply split into two
374 static ir_node *convert_dbl_to_int(ir_node *bl, ir_node *arg, ir_node *mem,
375 ir_node **resH, ir_node **resL) {
377 tarval *tv = get_Const_tarval(arg);
380 /* get the upper 32 bits */
381 v = get_tarval_sub_bits(tv, 7);
382 v = (v << 8) | get_tarval_sub_bits(tv, 6);
383 v = (v << 8) | get_tarval_sub_bits(tv, 5);
384 v = (v << 8) | get_tarval_sub_bits(tv, 4);
385 *resH = new_Const_long(mode_Is, v);
387 /* get the lower 32 bits */
388 v = get_tarval_sub_bits(tv, 3);
389 v = (v << 8) | get_tarval_sub_bits(tv, 2);
390 v = (v << 8) | get_tarval_sub_bits(tv, 1);
391 v = (v << 8) | get_tarval_sub_bits(tv, 0);
392 *resL = new_Const_long(mode_Is, v);
394 else if (get_irn_op(skip_Proj(arg)) == op_Load) {
395 /* FIXME: handling of low/high depends on LE/BE here */
399 ir_graph *irg = current_ir_graph;
402 conv = new_rd_arm_fpaDbl2GP(NULL, irg, bl, arg, mem);
404 *resL = new_r_Proj(irg, bl, conv, mode_Is, pn_arm_fpaDbl2GP_low);
405 *resH = new_r_Proj(irg, bl, conv, mode_Is, pn_arm_fpaDbl2GP_high);
406 mem = new_r_Proj(irg, bl, conv, mode_M, pn_arm_fpaDbl2GP_M);
412 * Move a single floating point value into an integer register.
413 * Place the move operation into block bl.
415 * Handle some special cases here:
416 * 1.) A constant: simply move
417 * 2.) A load: simply load
419 static ir_node *convert_sng_to_int(ir_node *bl, ir_node *arg)
424 tarval *tv = get_Const_tarval(arg);
427 /* get the lower 32 bits */
428 v = get_tarval_sub_bits(tv, 3);
429 v = (v << 8) | get_tarval_sub_bits(tv, 2);
430 v = (v << 8) | get_tarval_sub_bits(tv, 1);
431 v = (v << 8) | get_tarval_sub_bits(tv, 0);
432 return new_Const_long(mode_Is, v);
434 else if (get_irn_op(skip_Proj(arg)) == op_Load) {
437 load = skip_Proj(arg);
444 * Convert the arguments of a call to support the
445 * ARM calling convention of general purpose AND floating
448 static void handle_calls(ir_node *call, void *env)
450 arm_code_gen_t *cg = env;
451 int i, j, n, size, idx, flag, n_param, n_res, first_variadic;
452 ir_type *mtp, *new_mtd, *new_tp[5];
453 ir_node *new_in[5], **in;
459 /* check, if we need conversions */
460 n = get_Call_n_params(call);
461 mtp = get_Call_type(call);
462 assert(get_method_n_params(mtp) == n);
464 /* it's always enough to handle the first 4 parameters */
467 flag = size = idx = 0;
468 bl = get_nodes_block(call);
469 for (i = 0; i < n; ++i) {
470 ir_type *param_tp = get_method_param_type(mtp, i);
472 if (is_compound_type(param_tp)) {
473 /* an aggregate parameter: bad case */
477 /* a primitive parameter */
478 ir_mode *mode = get_type_mode(param_tp);
480 if (mode_is_float(mode)) {
481 if (get_mode_size_bits(mode) > 32) {
482 ir_node *mem = get_Call_mem(call);
484 /* Beware: ARM wants the high part first */
486 new_tp[idx] = cg->int_tp;
487 new_tp[idx+1] = cg->int_tp;
488 mem = convert_dbl_to_int(bl, get_Call_param(call, i), mem, &new_in[idx], &new_in[idx+1]);
490 set_Call_mem(call, mem);
494 new_tp[idx] = cg->int_tp;
495 new_in[idx] = convert_sng_to_int(bl, get_Call_param(call, i));
502 new_tp[idx] = param_tp;
503 new_in[idx] = get_Call_param(call, i);
512 /* if flag is NOT set, no need to translate the method type */
516 /* construct a new method type */
518 n_param = get_method_n_params(mtp) - n + idx;
519 n_res = get_method_n_ress(mtp);
520 new_mtd = new_d_type_method(get_type_ident(mtp), n_param, n_res, get_type_dbg_info(mtp));
522 for (i = 0; i < idx; ++i)
523 set_method_param_type(new_mtd, i, new_tp[i]);
524 for (i = n, j = idx; i < get_method_n_params(mtp); ++i)
525 set_method_param_type(new_mtd, j++, get_method_param_type(mtp, i));
526 for (i = 0; i < n_res; ++i)
527 set_method_res_type(new_mtd, i, get_method_res_type(mtp, i));
529 set_method_calling_convention(new_mtd, get_method_calling_convention(mtp));
530 first_variadic = get_method_first_variadic_param_index(mtp);
531 if (first_variadic >= 0)
532 set_method_first_variadic_param_index(new_mtd, first_variadic);
534 if (is_lowered_type(mtp)) {
535 mtp = get_associated_type(mtp);
537 set_lowered_type(mtp, new_mtd);
539 set_Call_type(call, new_mtd);
541 /* calculate new in array of the Call */
542 NEW_ARR_A(ir_node *, in, n_param + 2);
543 for (i = 0; i < idx; ++i)
544 in[2 + i] = new_in[i];
545 for (i = n, j = idx; i < get_method_n_params(mtp); ++i)
546 in[2 + j++] = get_Call_param(call, i);
548 in[0] = get_Call_mem(call);
549 in[1] = get_Call_ptr(call);
551 /* finally, change the call inputs */
552 set_irn_in(call, n_param + 2, in);
556 * Handle graph transformations before the abi converter does its work.
558 static void arm_before_abi(void *self) {
559 arm_code_gen_t *cg = self;
561 irg_walk_graph(cg->irg, NULL, handle_calls, cg);
565 static void *arm_cg_init(be_irg_t *birg);
567 static const arch_code_generator_if_t arm_code_gen_if = {
569 arm_before_abi, /* before abi introduce */
572 arm_before_sched, /* before scheduling hook */
573 arm_before_ra, /* before register allocation hook */
580 * Initializes the code generator.
582 static void *arm_cg_init(be_irg_t *birg) {
583 static ir_type *int_tp = NULL;
584 arm_isa_t *isa = (arm_isa_t *)birg->main_env->arch_env->isa;
588 /* create an integer type with machine size */
589 int_tp = new_type_primitive(new_id_from_chars("int", 3), mode_Is);
592 cg = xmalloc(sizeof(*cg));
593 cg->impl = &arm_code_gen_if;
595 cg->reg_set = new_set(arm_cmp_irn_reg_assoc, 1024);
596 cg->arch_env = birg->main_env->arch_env;
600 cg->have_fp_insn = 0;
601 cg->unknown_gp = NULL;
602 cg->unknown_fpa = NULL;
603 cg->dump = (birg->main_env->options->dump_flags & DUMP_BE) ? 1 : 0;
605 FIRM_DBG_REGISTER(cg->mod, "firm.be.arm.cg");
607 cur_reg_set = cg->reg_set;
611 /* enter the current code generator */
614 return (arch_code_generator_t *)cg;
619 * Maps all intrinsic calls that the backend support
620 * and map all instructions the backend did not support
623 static void arm_handle_intrinsics(void) {
624 ir_type *tp, *int_tp, *uint_tp;
628 #define ID(x) new_id_from_chars(x, sizeof(x)-1)
630 int_tp = new_type_primitive(ID("int"), mode_Is);
631 uint_tp = new_type_primitive(ID("uint"), mode_Iu);
633 /* ARM has neither a signed div instruction ... */
636 i_instr_record *map_Div = &records[n_records++].i_instr;
638 tp = new_type_method(ID("rt_iDiv"), 2, 1);
639 set_method_param_type(tp, 0, int_tp);
640 set_method_param_type(tp, 1, int_tp);
641 set_method_res_type(tp, 0, int_tp);
643 rt_Div.ent = new_entity(get_glob_type(), ID("__divsi3"), tp);
644 rt_Div.mode = mode_T;
645 rt_Div.res_mode = mode_Is;
646 rt_Div.mem_proj_nr = pn_Div_M;
647 rt_Div.regular_proj_nr = pn_Div_X_regular;
648 rt_Div.exc_proj_nr = pn_Div_X_except;
649 rt_Div.exc_mem_proj_nr = pn_Div_M;
650 rt_Div.res_proj_nr = pn_Div_res;
652 set_entity_visibility(rt_Div.ent, visibility_external_allocated);
654 map_Div->kind = INTRINSIC_INSTR;
655 map_Div->op = op_Div;
656 map_Div->i_mapper = (i_mapper_func)i_mapper_RuntimeCall;
657 map_Div->ctx = &rt_Div;
659 /* ... nor an unsigned div instruction ... */
662 i_instr_record *map_Div = &records[n_records++].i_instr;
664 tp = new_type_method(ID("rt_uDiv"), 2, 1);
665 set_method_param_type(tp, 0, uint_tp);
666 set_method_param_type(tp, 1, uint_tp);
667 set_method_res_type(tp, 0, uint_tp);
669 rt_Div.ent = new_entity(get_glob_type(), ID("__udivsi3"), tp);
670 rt_Div.mode = mode_T;
671 rt_Div.res_mode = mode_Iu;
672 rt_Div.mem_proj_nr = pn_Div_M;
673 rt_Div.regular_proj_nr = pn_Div_X_regular;
674 rt_Div.exc_proj_nr = pn_Div_X_except;
675 rt_Div.exc_mem_proj_nr = pn_Div_M;
676 rt_Div.res_proj_nr = pn_Div_res;
678 set_entity_visibility(rt_Div.ent, visibility_external_allocated);
680 map_Div->kind = INTRINSIC_INSTR;
681 map_Div->op = op_Div;
682 map_Div->i_mapper = (i_mapper_func)i_mapper_RuntimeCall;
683 map_Div->ctx = &rt_Div;
685 /* ... nor a signed mod instruction ... */
688 i_instr_record *map_Mod = &records[n_records++].i_instr;
690 tp = new_type_method(ID("rt_iMod"), 2, 1);
691 set_method_param_type(tp, 0, int_tp);
692 set_method_param_type(tp, 1, int_tp);
693 set_method_res_type(tp, 0, int_tp);
695 rt_Mod.ent = new_entity(get_glob_type(), ID("__modsi3"), tp);
696 rt_Mod.mode = mode_T;
697 rt_Mod.res_mode = mode_Is;
698 rt_Mod.mem_proj_nr = pn_Mod_M;
699 rt_Mod.regular_proj_nr = pn_Mod_X_regular;
700 rt_Mod.exc_proj_nr = pn_Mod_X_except;
701 rt_Mod.exc_mem_proj_nr = pn_Mod_M;
702 rt_Mod.res_proj_nr = pn_Mod_res;
704 set_entity_visibility(rt_Mod.ent, visibility_external_allocated);
706 map_Mod->kind = INTRINSIC_INSTR;
707 map_Mod->op = op_Mod;
708 map_Mod->i_mapper = (i_mapper_func)i_mapper_RuntimeCall;
709 map_Mod->ctx = &rt_Mod;
711 /* ... nor an unsigned mod. */
714 i_instr_record *map_Mod = &records[n_records++].i_instr;
716 tp = new_type_method(ID("rt_uMod"), 2, 1);
717 set_method_param_type(tp, 0, uint_tp);
718 set_method_param_type(tp, 1, uint_tp);
719 set_method_res_type(tp, 0, uint_tp);
721 rt_Mod.ent = new_entity(get_glob_type(), ID("__umodsi3"), tp);
722 rt_Mod.mode = mode_T;
723 rt_Mod.res_mode = mode_Iu;
724 rt_Mod.mem_proj_nr = pn_Mod_M;
725 rt_Mod.regular_proj_nr = pn_Mod_X_regular;
726 rt_Mod.exc_proj_nr = pn_Mod_X_except;
727 rt_Mod.exc_mem_proj_nr = pn_Mod_M;
728 rt_Mod.res_proj_nr = pn_Mod_res;
730 set_entity_visibility(rt_Mod.ent, visibility_external_allocated);
732 map_Mod->kind = INTRINSIC_INSTR;
733 map_Mod->op = op_Mod;
734 map_Mod->i_mapper = (i_mapper_func)i_mapper_RuntimeCall;
735 map_Mod->ctx = &rt_Mod;
739 lower_intrinsics(records, n_records);
742 /*****************************************************************
743 * ____ _ _ _____ _____
744 * | _ \ | | | | |_ _|/ ____| /\
745 * | |_) | __ _ ___| | _____ _ __ __| | | | | (___ / \
746 * | _ < / _` |/ __| |/ / _ \ '_ \ / _` | | | \___ \ / /\ \
747 * | |_) | (_| | (__| < __/ | | | (_| | _| |_ ____) / ____ \
748 * |____/ \__,_|\___|_|\_\___|_| |_|\__,_| |_____|_____/_/ \_\
750 *****************************************************************/
752 static arm_isa_t arm_isa_template = {
754 &arm_isa_if, /* isa interface */
755 &arm_gp_regs[REG_SP], /* stack pointer */
756 &arm_gp_regs[REG_R11], /* base pointer */
757 -1, /* stack direction */
758 NULL, /* main environment */
760 5, /* reload costs */
762 0, /* use generic register names instead of SP, LR, PC */
763 ARM_FPU_ARCH_FPE, /* FPU architecture */
764 NULL, /* current code generator */
765 NULL_EMITTER, /* emitter environment */
769 * Initializes the backend ISA and opens the output file.
771 static void *arm_init(FILE *file_handle) {
772 static int inited = 0;
778 isa = xmalloc(sizeof(*isa));
779 memcpy(isa, &arm_isa_template, sizeof(*isa));
784 be_emit_init_env(&isa->emit, file_handle);
786 arm_create_opcodes();
787 arm_register_copy_attr_func();
788 arm_handle_intrinsics();
790 /* we mark referenced global entities, so we can only emit those which
791 * are actually referenced. (Note: you mustn't use the type visited flag
792 * elsewhere in the backend)
794 inc_master_type_visited();
803 * Closes the output file and frees the ISA structure.
805 static void arm_done(void *self) {
806 arm_isa_t *isa = self;
808 be_gas_emit_decls(&isa->emit, isa->arch_isa.main_env, 1);
810 be_emit_destroy_env(&isa->emit);
816 * Report the number of register classes.
817 * If we don't have fp instructions, report only GP
818 * here to speed up register allocation (and makes dumps
819 * smaller and more readable).
821 static int arm_get_n_reg_class(const void *self) {
822 const arm_isa_t *isa = self;
824 /* ARGH! is called BEFORE transform */
826 return isa->cg->have_fp_insn ? 2 : 1;
830 * Return the register class with requested index.
832 static const arch_register_class_t *arm_get_reg_class(const void *self, int i) {
834 return i == 0 ? &arm_reg_classes[CLASS_arm_gp] : &arm_reg_classes[CLASS_arm_fpa];
838 * Get the register class which shall be used to store a value of a given mode.
839 * @param self The this pointer.
840 * @param mode The mode in question.
841 * @return A register class which can hold values of the given mode.
843 const arch_register_class_t *arm_get_reg_class_for_mode(const void *self, const ir_mode *mode) {
845 if (mode_is_float(mode))
846 return &arm_reg_classes[CLASS_arm_fpa];
848 return &arm_reg_classes[CLASS_arm_gp];
852 * Produces the type which sits between the stack args and the locals on the stack.
853 * it will contain the return address and space to store the old base pointer.
854 * @return The Firm type modelling the ABI between type.
856 static ir_type *arm_get_between_type(void *self) {
857 static ir_type *between_type = NULL;
858 static ir_entity *old_bp_ent = NULL;
862 ir_entity *ret_addr_ent;
863 ir_type *ret_addr_type = new_type_primitive(new_id_from_str("return_addr"), mode_P);
864 ir_type *old_bp_type = new_type_primitive(new_id_from_str("bp"), mode_P);
866 between_type = new_type_class(new_id_from_str("arm_between_type"));
867 old_bp_ent = new_entity(between_type, new_id_from_str("old_bp"), old_bp_type);
868 ret_addr_ent = new_entity(between_type, new_id_from_str("old_bp"), ret_addr_type);
870 set_entity_offset(old_bp_ent, 0);
871 set_entity_offset(ret_addr_ent, get_type_size_bytes(old_bp_type));
872 set_type_size_bytes(between_type, get_type_size_bytes(old_bp_type) + get_type_size_bytes(ret_addr_type));
880 be_abi_call_flags_bits_t flags;
881 const arch_env_t *arch_env;
882 const arch_isa_t *isa;
886 static void *arm_abi_init(const be_abi_call_t *call, const arch_env_t *arch_env, ir_graph *irg)
888 arm_abi_env_t *env = xmalloc(sizeof(env[0]));
889 be_abi_call_flags_t fl = be_abi_call_get_flags(call);
890 env->flags = fl.bits;
892 env->arch_env = arch_env;
893 env->isa = arch_env->isa;
897 static void arm_abi_dont_save_regs(void *self, pset *s)
899 arm_abi_env_t *env = self;
900 if (env->flags.try_omit_fp)
901 pset_insert_ptr(s, env->isa->bp);
907 * Build the ARM prolog
909 static const arch_register_t *arm_abi_prologue(void *self, ir_node **mem, pmap *reg_map) {
910 ir_node *keep, *store;
911 arm_abi_env_t *env = self;
912 ir_graph *irg = env->irg;
913 ir_node *block = get_irg_start_block(irg);
914 arch_register_class_t *gp = &arm_reg_classes[CLASS_arm_gp];
916 ir_node *fp = be_abi_reg_map_get(reg_map, env->isa->bp);
917 ir_node *ip = be_abi_reg_map_get(reg_map, &arm_gp_regs[REG_R12]);
918 ir_node *sp = be_abi_reg_map_get(reg_map, env->isa->sp);
919 ir_node *lr = be_abi_reg_map_get(reg_map, &arm_gp_regs[REG_LR]);
920 ir_node *pc = be_abi_reg_map_get(reg_map, &arm_gp_regs[REG_PC]);
922 if(env->flags.try_omit_fp)
925 ip = be_new_Copy(gp, irg, block, sp );
926 arch_set_irn_register(env->arch_env, ip, &arm_gp_regs[REG_R12]);
927 be_set_constr_single_reg(ip, BE_OUT_POS(0), &arm_gp_regs[REG_R12] );
929 store = new_rd_arm_StoreStackM4Inc(NULL, irg, block, sp, fp, ip, lr, pc, *mem);
931 sp = new_r_Proj(irg, block, store, env->isa->sp->reg_class->mode, pn_arm_StoreStackM4Inc_ptr);
932 arch_set_irn_register(env->arch_env, sp, env->isa->sp);
933 *mem = new_r_Proj(irg, block, store, mode_M, pn_arm_StoreStackM4Inc_M);
935 keep = be_new_CopyKeep_single(gp, irg, block, ip, sp, get_irn_mode(ip));
936 be_node_set_reg_class(keep, 1, gp);
937 arch_set_irn_register(env->arch_env, keep, &arm_gp_regs[REG_R12]);
938 be_set_constr_single_reg(keep, BE_OUT_POS(0), &arm_gp_regs[REG_R12] );
940 fp = new_rd_arm_Sub_i(NULL, irg, block, keep, get_irn_mode(fp),
941 new_tarval_from_long(4, get_irn_mode(fp)));
942 arch_set_irn_register(env->arch_env, fp, env->isa->bp);
944 be_abi_reg_map_set(reg_map, env->isa->bp, fp);
945 be_abi_reg_map_set(reg_map, &arm_gp_regs[REG_R12], keep);
946 be_abi_reg_map_set(reg_map, env->isa->sp, sp);
947 be_abi_reg_map_set(reg_map, &arm_gp_regs[REG_LR], lr);
948 be_abi_reg_map_set(reg_map, &arm_gp_regs[REG_PC], pc);
953 static void arm_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_map) {
954 arm_abi_env_t *env = self;
955 ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp);
956 ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp);
957 ir_node *curr_pc = be_abi_reg_map_get(reg_map, &arm_gp_regs[REG_PC]);
958 ir_node *curr_lr = be_abi_reg_map_get(reg_map, &arm_gp_regs[REG_LR]);
960 // TODO: Activate Omit fp in epilogue
961 if(env->flags.try_omit_fp) {
962 curr_sp = be_new_IncSP(env->isa->sp, env->irg, bl, curr_sp, BE_STACK_FRAME_SIZE_SHRINK);
963 add_irn_dep(curr_sp, *mem);
965 curr_lr = be_new_CopyKeep_single(&arm_reg_classes[CLASS_arm_gp], env->irg, bl, curr_lr, curr_sp, get_irn_mode(curr_lr));
966 be_node_set_reg_class(curr_lr, 1, &arm_reg_classes[CLASS_arm_gp]);
967 arch_set_irn_register(env->arch_env, curr_lr, &arm_gp_regs[REG_LR]);
968 be_set_constr_single_reg(curr_lr, BE_OUT_POS(0), &arm_gp_regs[REG_LR] );
970 curr_pc = be_new_Copy(&arm_reg_classes[CLASS_arm_gp], env->irg, bl, curr_lr );
971 arch_set_irn_register(env->arch_env, curr_pc, &arm_gp_regs[REG_PC]);
972 be_set_constr_single_reg(curr_pc, BE_OUT_POS(0), &arm_gp_regs[REG_PC] );
973 be_node_set_flags(curr_pc, BE_OUT_POS(0), arch_irn_flags_ignore);
977 tarval *tv = new_tarval_from_long(12,mode_Iu);
978 sub12_node = new_rd_arm_Sub_i(NULL, env->irg, bl, curr_bp, mode_Iu, tv);
980 //set_arm_req_out_all(sub12_node, sub12_req);
981 arch_set_irn_register(env->arch_env, sub12_node, env->isa->sp);
982 load_node = new_rd_arm_LoadStackM3( NULL, env->irg, bl, sub12_node, *mem );
984 //set_arm_req_out(load_node, &arm_default_req_arm_gp_r11, 0);
985 //set_arm_req_out(load_node, &arm_default_req_arm_gp_sp, 1);
986 //set_arm_req_out(load_node, &arm_default_req_arm_gp_pc, 2);
987 curr_bp = new_r_Proj(env->irg, bl, load_node, env->isa->bp->reg_class->mode, pn_arm_LoadStackM3_res0);
988 curr_sp = new_r_Proj(env->irg, bl, load_node, env->isa->sp->reg_class->mode, pn_arm_LoadStackM3_res1);
989 curr_pc = new_r_Proj(env->irg, bl, load_node, mode_Iu, pn_arm_LoadStackM3_res2);
990 *mem = new_r_Proj(env->irg, bl, load_node, mode_M, pn_arm_LoadStackM3_M);
991 arch_set_irn_register(env->arch_env, curr_bp, env->isa->bp);
992 arch_set_irn_register(env->arch_env, curr_sp, env->isa->sp);
993 arch_set_irn_register(env->arch_env, curr_pc, &arm_gp_regs[REG_PC]);
995 be_abi_reg_map_set(reg_map, env->isa->sp, curr_sp);
996 be_abi_reg_map_set(reg_map, env->isa->bp, curr_bp);
997 be_abi_reg_map_set(reg_map, &arm_gp_regs[REG_LR], curr_lr);
998 be_abi_reg_map_set(reg_map, &arm_gp_regs[REG_PC], curr_pc);
1001 static const be_abi_callbacks_t arm_abi_callbacks = {
1004 arm_get_between_type,
1005 arm_abi_dont_save_regs,
1012 * Get the ABI restrictions for procedure calls.
1013 * @param self The this pointer.
1014 * @param method_type The type of the method (procedure) in question.
1015 * @param abi The abi object to be modified
1017 void arm_get_call_abi(const void *self, ir_type *method_type, be_abi_call_t *abi) {
1021 int n = get_method_n_params(method_type);
1022 be_abi_call_flags_t flags = {
1024 0, /* store from left to right */
1025 0, /* store arguments sequential */
1026 1, /* try to omit the frame pointer */
1027 1, /* the function can use any register as frame pointer */
1028 1, /* a call can take the callee's address as an immediate */
1029 0, /* IRG is a leaf function */
1030 0 /* Set to one, if there is already enough room on the stack for call args. */
1035 /* set stack parameter passing style */
1036 be_abi_call_set_flags(abi, flags, &arm_abi_callbacks);
1038 for (i = 0; i < n; i++) {
1039 /* reg = get reg for param i; */
1040 /* be_abi_call_param_reg(abi, i, reg); */
1043 be_abi_call_param_reg(abi, i, arm_get_RegParam_reg(i));
1045 be_abi_call_param_stack(abi, i, 4, 0, 0);
1048 /* set return registers */
1049 n = get_method_n_ress(method_type);
1051 assert(n <= 2 && "more than two results not supported");
1053 /* In case of 64bit returns, we will have two 32bit values */
1055 tp = get_method_res_type(method_type, 0);
1056 mode = get_type_mode(tp);
1058 assert(!mode_is_float(mode) && "two FP results not supported");
1060 tp = get_method_res_type(method_type, 1);
1061 mode = get_type_mode(tp);
1063 assert(!mode_is_float(mode) && "mixed INT, FP results not supported");
1065 be_abi_call_res_reg(abi, 0, &arm_gp_regs[REG_R0]);
1066 be_abi_call_res_reg(abi, 1, &arm_gp_regs[REG_R1]);
1067 } else if (n == 1) {
1068 const arch_register_t *reg;
1070 tp = get_method_res_type(method_type, 0);
1071 assert(is_atomic_type(tp));
1072 mode = get_type_mode(tp);
1074 reg = mode_is_float(mode) ? &arm_fpa_regs[REG_F0] : &arm_gp_regs[REG_R0];
1075 be_abi_call_res_reg(abi, 0, reg);
1079 static const void *arm_get_irn_ops(const arch_irn_handler_t *self, const ir_node *irn) {
1082 return &arm_irn_ops;
1085 const arch_irn_handler_t arm_irn_handler = {
1089 const arch_irn_handler_t *arm_get_irn_handler(const void *self) {
1091 return &arm_irn_handler;
1094 int arm_to_appear_in_schedule(void *block_env, const ir_node *irn) {
1096 if(!is_arm_irn(irn))
1103 * Initializes the code generator interface.
1105 static const arch_code_generator_if_t *arm_get_code_generator_if(void *self) {
1107 return &arm_code_gen_if;
1110 list_sched_selector_t arm_sched_selector;
1113 * Returns the reg_pressure scheduler with to_appear_in_schedule() over\loaded
1115 static const list_sched_selector_t *arm_get_list_sched_selector(const void *self, list_sched_selector_t *selector) {
1118 memcpy(&arm_sched_selector, reg_pressure_selector, sizeof(list_sched_selector_t));
1119 arm_sched_selector.to_appear_in_schedule = arm_to_appear_in_schedule;
1120 return &arm_sched_selector;
1123 static const ilp_sched_selector_t *arm_get_ilp_sched_selector(const void *self) {
1129 * Returns the necessary byte alignment for storing a register of given class.
1131 static int arm_get_reg_class_alignment(const void *self, const arch_register_class_t *cls) {
1134 /* ARM is a 32 bit CPU, no need for other alignment */
1138 static const be_execution_unit_t ***arm_get_allowed_execution_units(const void *self, const ir_node *irn) {
1146 static const be_machine_t *arm_get_machine(const void *self) {
1154 * Return irp irgs in the desired order.
1156 static ir_graph **arm_get_irg_list(const void *self, ir_graph ***irg_list) {
1163 * Returns the libFirm configuration parameter for this backend.
1165 static const backend_params *arm_get_libfirm_params(void) {
1166 static arch_dep_params_t ad = {
1168 1, /* Muls are fast enough on ARM but ... */
1169 1, /* ... one shift would be possible better */
1170 0, /* SMUL is needed, only in Arch M*/
1171 0, /* UMUL is needed, only in Arch M */
1172 32, /* SMUL & UMUL available for 32 bit */
1174 static backend_params p = {
1175 1, /* need dword lowering */
1176 0, /* don't support inline assembler yet */
1177 NULL, /* no additional opcodes */
1178 NULL, /* will be set later */
1179 NULL, /* but yet no creator function */
1180 NULL, /* context for create_intrinsic_fkt */
1181 NULL, /* no if conversion settings */
1188 /* fpu set architectures. */
1189 static const lc_opt_enum_int_items_t arm_fpu_items[] = {
1190 { "softfloat", ARM_FPU_ARCH_SOFTFLOAT },
1191 { "fpe", ARM_FPU_ARCH_FPE },
1192 { "fpa", ARM_FPU_ARCH_FPA },
1193 { "vfp1xd", ARM_FPU_ARCH_VFP_V1xD },
1194 { "vfp1", ARM_FPU_ARCH_VFP_V1 },
1195 { "vfp2", ARM_FPU_ARCH_VFP_V2 },
1199 static lc_opt_enum_int_var_t arch_fpu_var = {
1200 &arm_isa_template.fpu_arch, arm_fpu_items
1203 static const lc_opt_table_entry_t arm_options[] = {
1204 LC_OPT_ENT_ENUM_INT("fpunit", "select the floating point unit", &arch_fpu_var),
1205 LC_OPT_ENT_BOOL("gen_reg_names", "use generic register names", &arm_isa_template.gen_reg_names),
1209 const arch_isa_if_t arm_isa_if = {
1212 arm_get_n_reg_class,
1214 arm_get_reg_class_for_mode,
1216 arm_get_irn_handler,
1217 arm_get_code_generator_if,
1218 arm_get_list_sched_selector,
1219 arm_get_ilp_sched_selector,
1220 arm_get_reg_class_alignment,
1221 arm_get_libfirm_params,
1222 arm_get_allowed_execution_units,
1227 void be_init_arch_arm(void)
1229 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
1230 lc_opt_entry_t *arm_grp = lc_opt_get_grp(be_grp, "arm");
1232 lc_opt_add_table(arm_grp, arm_options);
1234 be_register_isa_if("arm", &arm_isa_if);
1236 arm_init_transform();
1239 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_arch_arm);