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"
46 #include "../bearch_t.h" /* the general register allocator interface */
47 #include "../benode_t.h"
48 #include "../belower.h"
49 #include "../besched_t.h"
52 #include "../bemachine.h"
53 #include "../beilpsched.h"
54 #include "../bemodule.h"
55 #include "../beirg_t.h"
56 #include "../bespillslots.h"
57 #include "../begnuas.h"
59 #include "bearch_arm_t.h"
61 #include "arm_new_nodes.h" /* arm nodes interface */
62 #include "gen_arm_regalloc_if.h" /* the generated interface (register type and class defenitions) */
63 #include "arm_transform.h"
64 #include "arm_emitter.h"
65 #include "arm_map_regs.h"
67 #define DEBUG_MODULE "firm.be.arm.isa"
69 /* TODO: ugly, but we need it to get access to the registers assigned to Phi nodes */
70 static set *cur_reg_set = NULL;
72 /**************************************************
75 * _ __ ___ __ _ __ _| | | ___ ___ _| |_
76 * | '__/ _ \/ _` | / _` | | |/ _ \ / __| | | _|
77 * | | | __/ (_| | | (_| | | | (_) | (__ | | |
78 * |_| \___|\__, | \__,_|_|_|\___/ \___| |_|_|
81 **************************************************/
84 * Return register requirements for a arm node.
85 * If the node returns a tuple (mode_T) then the proj's
86 * will be asked for this information.
89 arch_register_req_t *arm_get_irn_reg_req(const void *self, const ir_node *node,
92 long node_pos = pos == -1 ? 0 : pos;
93 ir_mode *mode = get_irn_mode(node);
96 if (is_Block(node) || mode == mode_X) {
97 return arch_no_register_req;
100 if (mode == mode_T && pos < 0) {
101 return arch_no_register_req;
106 return arch_no_register_req;
109 return arch_no_register_req;
112 node_pos = (pos == -1) ? get_Proj_proj(node) : pos;
113 node = skip_Proj_const(node);
116 /* get requirements for our own nodes */
117 if (is_arm_irn(node)) {
118 const arch_register_req_t *req;
120 req = get_arm_in_req(node, pos);
122 req = get_arm_out_req(node, node_pos);
128 /* unknown should be transformed by now */
129 assert(!is_Unknown(node));
130 return arch_no_register_req;
133 static void arm_set_irn_reg(const void *self, ir_node *irn,
134 const arch_register_t *reg)
139 if (get_irn_mode(irn) == mode_X) {
144 pos = get_Proj_proj(irn);
145 irn = skip_Proj(irn);
148 if (is_arm_irn(irn)) {
149 const arch_register_t **slots;
151 slots = get_arm_slots(irn);
155 /* here we set the registers for the Phi nodes */
156 arm_set_firm_reg(irn, reg, cur_reg_set);
160 static const arch_register_t *arm_get_irn_reg(const void *self,
164 const arch_register_t *reg = NULL;
169 if (get_irn_mode(irn) == mode_X) {
173 pos = get_Proj_proj(irn);
174 irn = skip_Proj_const(irn);
177 if (is_arm_irn(irn)) {
178 const arch_register_t **slots;
179 slots = get_arm_slots(irn);
183 reg = arm_get_firm_reg(irn, cur_reg_set);
189 static arch_irn_class_t arm_classify(const void *self, const ir_node *irn)
192 irn = skip_Proj_const(irn);
195 return arch_irn_class_branch;
197 else if (is_arm_irn(irn)) {
198 return arch_irn_class_normal;
204 static arch_irn_flags_t arm_get_flags(const void *self, const ir_node *irn)
207 irn = skip_Proj_const(irn);
209 if (is_arm_irn(irn)) {
210 return get_arm_flags(irn);
212 else if (is_Unknown(irn)) {
213 return arch_irn_flags_ignore;
219 static ir_entity *arm_get_frame_entity(const void *self, const ir_node *irn)
223 /* TODO: return the entity assigned to the frame */
227 static void arm_set_frame_entity(const void *self, ir_node *irn, ir_entity *ent)
232 /* TODO: set the entity assigned to the frame */
236 * This function is called by the generic backend to correct offsets for
237 * nodes accessing the stack.
239 static void arm_set_stack_bias(const void *self, ir_node *irn, int bias)
244 /* TODO: correct offset if irn accesses the stack */
247 static int arm_get_sp_bias(const void *self, const ir_node *irn)
254 /* fill register allocator interface */
256 static const arch_irn_ops_if_t arm_irn_ops_if = {
262 arm_get_frame_entity,
263 arm_set_frame_entity,
266 NULL, /* get_inverse */
267 NULL, /* get_op_estimated_cost */
268 NULL, /* possible_memory_operand */
269 NULL, /* perform_memory_operand */
272 arm_irn_ops_t arm_irn_ops = {
279 /**************************************************
282 * ___ ___ __| | ___ __ _ ___ _ __ _| |_
283 * / __/ _ \ / _` |/ _ \/ _` |/ _ \ '_ \ | | _|
284 * | (_| (_) | (_| | __/ (_| | __/ | | | | | |
285 * \___\___/ \__,_|\___|\__, |\___|_| |_| |_|_|
288 **************************************************/
291 * Transforms the standard Firm graph into
294 static void arm_prepare_graph(void *self) {
295 arm_code_gen_t *cg = self;
297 /* transform nodes into assembler instructions */
298 arm_transform_graph(cg);
300 /* do local optimizations (mainly CSE) */
301 local_optimize_graph(cg->irg);
304 be_dump(cg->irg, "-transformed", dump_ir_block_graph_sched);
306 /* do code placement, to optimize the position of constants */
310 be_dump(cg->irg, "-place", dump_ir_block_graph_sched);
314 * Called immediately before emit phase.
316 static void arm_finish_irg(void *self)
319 /* TODO: - fix offsets for nodes accessing stack
326 * These are some hooks which must be filled but are probably not needed.
328 static void arm_before_sched(void *self)
331 /* Some stuff you need to do after scheduling but before register allocation */
334 static void arm_before_ra(void *self)
337 /* Some stuff you need to do immediately after register allocation */
341 * We transform Spill and Reload here. This needs to be done before
342 * stack biasing otherwise we would miss the corrected offset for these nodes.
344 static void arm_after_ra(void *self)
346 arm_code_gen_t *cg = self;
347 be_coalesce_spillslots(cg->birg);
351 * Emits the code, closes the output file and frees
352 * the code generator interface.
354 static void arm_emit_and_done(void *self) {
355 arm_code_gen_t *cg = self;
356 ir_graph *irg = cg->irg;
358 arm_gen_routine(cg, irg);
362 /* de-allocate code generator */
363 del_set(cg->reg_set);
368 * Move a double floating point value into an integer register.
369 * Place the move operation into block bl.
371 * Handle some special cases here:
372 * 1.) A constant: simply split into two
373 * 2.) A load: simply split into two
375 static ir_node *convert_dbl_to_int(ir_node *bl, ir_node *arg, ir_node *mem,
376 ir_node **resH, ir_node **resL) {
378 tarval *tv = get_Const_tarval(arg);
381 /* get the upper 32 bits */
382 v = get_tarval_sub_bits(tv, 7);
383 v = (v << 8) | get_tarval_sub_bits(tv, 6);
384 v = (v << 8) | get_tarval_sub_bits(tv, 5);
385 v = (v << 8) | get_tarval_sub_bits(tv, 4);
386 *resH = new_Const_long(mode_Is, v);
388 /* get the lower 32 bits */
389 v = get_tarval_sub_bits(tv, 3);
390 v = (v << 8) | get_tarval_sub_bits(tv, 2);
391 v = (v << 8) | get_tarval_sub_bits(tv, 1);
392 v = (v << 8) | get_tarval_sub_bits(tv, 0);
393 *resL = new_Const_long(mode_Is, v);
395 else if (get_irn_op(skip_Proj(arg)) == op_Load) {
396 /* FIXME: handling of low/high depends on LE/BE here */
400 ir_graph *irg = current_ir_graph;
403 conv = new_rd_arm_fpaDbl2GP(NULL, irg, bl, arg, mem);
405 *resL = new_r_Proj(irg, bl, conv, mode_Is, pn_arm_fpaDbl2GP_low);
406 *resH = new_r_Proj(irg, bl, conv, mode_Is, pn_arm_fpaDbl2GP_high);
407 mem = new_r_Proj(irg, bl, conv, mode_M, pn_arm_fpaDbl2GP_M);
413 * Move a single floating point value into an integer register.
414 * Place the move operation into block bl.
416 * Handle some special cases here:
417 * 1.) A constant: simply move
418 * 2.) A load: simply load
420 static ir_node *convert_sng_to_int(ir_node *bl, ir_node *arg)
425 tarval *tv = get_Const_tarval(arg);
428 /* get the lower 32 bits */
429 v = get_tarval_sub_bits(tv, 3);
430 v = (v << 8) | get_tarval_sub_bits(tv, 2);
431 v = (v << 8) | get_tarval_sub_bits(tv, 1);
432 v = (v << 8) | get_tarval_sub_bits(tv, 0);
433 return new_Const_long(mode_Is, v);
435 else if (get_irn_op(skip_Proj(arg)) == op_Load) {
438 load = skip_Proj(arg);
445 * Convert the arguments of a call to support the
446 * ARM calling convention of general purpose AND floating
449 static void handle_calls(ir_node *call, void *env)
451 arm_code_gen_t *cg = env;
452 int i, j, n, size, idx, flag, n_param, n_res, first_variadic;
453 ir_type *mtp, *new_mtd, *new_tp[5];
454 ir_node *new_in[5], **in;
460 /* check, if we need conversions */
461 n = get_Call_n_params(call);
462 mtp = get_Call_type(call);
463 assert(get_method_n_params(mtp) == n);
465 /* it's always enough to handle the first 4 parameters */
468 flag = size = idx = 0;
469 bl = get_nodes_block(call);
470 for (i = 0; i < n; ++i) {
471 ir_type *param_tp = get_method_param_type(mtp, i);
473 if (is_compound_type(param_tp)) {
474 /* an aggregate parameter: bad case */
478 /* a primitive parameter */
479 ir_mode *mode = get_type_mode(param_tp);
481 if (mode_is_float(mode)) {
482 if (get_mode_size_bits(mode) > 32) {
483 ir_node *mem = get_Call_mem(call);
485 /* Beware: ARM wants the high part first */
487 new_tp[idx] = cg->int_tp;
488 new_tp[idx+1] = cg->int_tp;
489 mem = convert_dbl_to_int(bl, get_Call_param(call, i), mem, &new_in[idx], &new_in[idx+1]);
491 set_Call_mem(call, mem);
495 new_tp[idx] = cg->int_tp;
496 new_in[idx] = convert_sng_to_int(bl, get_Call_param(call, i));
503 new_tp[idx] = param_tp;
504 new_in[idx] = get_Call_param(call, i);
513 /* if flag is NOT set, no need to translate the method type */
517 /* construct a new method type */
519 n_param = get_method_n_params(mtp) - n + idx;
520 n_res = get_method_n_ress(mtp);
521 new_mtd = new_d_type_method(get_type_ident(mtp), n_param, n_res, get_type_dbg_info(mtp));
523 for (i = 0; i < idx; ++i)
524 set_method_param_type(new_mtd, i, new_tp[i]);
525 for (i = n, j = idx; i < get_method_n_params(mtp); ++i)
526 set_method_param_type(new_mtd, j++, get_method_param_type(mtp, i));
527 for (i = 0; i < n_res; ++i)
528 set_method_res_type(new_mtd, i, get_method_res_type(mtp, i));
530 set_method_calling_convention(new_mtd, get_method_calling_convention(mtp));
531 first_variadic = get_method_first_variadic_param_index(mtp);
532 if (first_variadic >= 0)
533 set_method_first_variadic_param_index(new_mtd, first_variadic);
535 if (is_lowered_type(mtp)) {
536 mtp = get_associated_type(mtp);
538 set_lowered_type(mtp, new_mtd);
540 set_Call_type(call, new_mtd);
542 /* calculate new in array of the Call */
543 NEW_ARR_A(ir_node *, in, n_param + 2);
544 for (i = 0; i < idx; ++i)
545 in[2 + i] = new_in[i];
546 for (i = n, j = idx; i < get_method_n_params(mtp); ++i)
547 in[2 + j++] = get_Call_param(call, i);
549 in[0] = get_Call_mem(call);
550 in[1] = get_Call_ptr(call);
552 /* finally, change the call inputs */
553 set_irn_in(call, n_param + 2, in);
557 * Handle graph transformations before the abi converter does its work.
559 static void arm_before_abi(void *self) {
560 arm_code_gen_t *cg = self;
562 irg_walk_graph(cg->irg, NULL, handle_calls, cg);
566 static void *arm_cg_init(be_irg_t *birg);
568 static const arch_code_generator_if_t arm_code_gen_if = {
570 arm_before_abi, /* before abi introduce */
573 arm_before_sched, /* before scheduling hook */
574 arm_before_ra, /* before register allocation hook */
581 * Initializes the code generator.
583 static void *arm_cg_init(be_irg_t *birg) {
584 static ir_type *int_tp = NULL;
585 arm_isa_t *isa = (arm_isa_t *)birg->main_env->arch_env->isa;
589 /* create an integer type with machine size */
590 int_tp = new_type_primitive(new_id_from_chars("int", 3), mode_Is);
593 cg = xmalloc(sizeof(*cg));
594 cg->impl = &arm_code_gen_if;
596 cg->reg_set = new_set(arm_cmp_irn_reg_assoc, 1024);
597 cg->arch_env = birg->main_env->arch_env;
601 cg->have_fp_insn = 0;
602 cg->unknown_gp = NULL;
603 cg->unknown_fpa = NULL;
604 cg->dump = (birg->main_env->options->dump_flags & DUMP_BE) ? 1 : 0;
606 FIRM_DBG_REGISTER(cg->mod, "firm.be.arm.cg");
608 cur_reg_set = cg->reg_set;
612 /* enter the current code generator */
615 return (arch_code_generator_t *)cg;
620 * Maps all intrinsic calls that the backend support
621 * and map all instructions the backend did not support
624 static void arm_handle_intrinsics(void) {
625 ir_type *tp, *int_tp, *uint_tp;
629 runtime_rt rt_iDiv, rt_uDiv, rt_iMod, rt_uMod;
631 #define ID(x) new_id_from_chars(x, sizeof(x)-1)
633 int_tp = new_type_primitive(ID("int"), mode_Is);
634 uint_tp = new_type_primitive(ID("uint"), mode_Iu);
636 /* ARM has neither a signed div instruction ... */
638 i_instr_record *map_Div = &records[n_records++].i_instr;
640 tp = new_type_method(ID("rt_iDiv"), 2, 1);
641 set_method_param_type(tp, 0, int_tp);
642 set_method_param_type(tp, 1, int_tp);
643 set_method_res_type(tp, 0, int_tp);
645 rt_iDiv.ent = new_entity(get_glob_type(), ID("__divsi3"), tp);
646 set_entity_ld_ident(rt_iDiv.ent, ID("__divsi3"));
647 rt_iDiv.mode = mode_T;
648 rt_iDiv.res_mode = mode_Is;
649 rt_iDiv.mem_proj_nr = pn_Div_M;
650 rt_iDiv.regular_proj_nr = pn_Div_X_regular;
651 rt_iDiv.exc_proj_nr = pn_Div_X_except;
652 rt_iDiv.exc_mem_proj_nr = pn_Div_M;
653 rt_iDiv.res_proj_nr = pn_Div_res;
655 set_entity_visibility(rt_iDiv.ent, visibility_external_allocated);
657 map_Div->kind = INTRINSIC_INSTR;
658 map_Div->op = op_Div;
659 map_Div->i_mapper = (i_mapper_func)i_mapper_RuntimeCall;
660 map_Div->ctx = &rt_iDiv;
662 /* ... nor an unsigned div instruction ... */
664 i_instr_record *map_Div = &records[n_records++].i_instr;
666 tp = new_type_method(ID("rt_uDiv"), 2, 1);
667 set_method_param_type(tp, 0, uint_tp);
668 set_method_param_type(tp, 1, uint_tp);
669 set_method_res_type(tp, 0, uint_tp);
671 rt_uDiv.ent = new_entity(get_glob_type(), ID("__udivsi3"), tp);
672 set_entity_ld_ident(rt_uDiv.ent, ID("__udivsi3"));
673 rt_uDiv.mode = mode_T;
674 rt_uDiv.res_mode = mode_Iu;
675 rt_uDiv.mem_proj_nr = pn_Div_M;
676 rt_uDiv.regular_proj_nr = pn_Div_X_regular;
677 rt_uDiv.exc_proj_nr = pn_Div_X_except;
678 rt_uDiv.exc_mem_proj_nr = pn_Div_M;
679 rt_uDiv.res_proj_nr = pn_Div_res;
681 set_entity_visibility(rt_uDiv.ent, visibility_external_allocated);
683 map_Div->kind = INTRINSIC_INSTR;
684 map_Div->op = op_Div;
685 map_Div->i_mapper = (i_mapper_func)i_mapper_RuntimeCall;
686 map_Div->ctx = &rt_uDiv;
688 /* ... nor a signed mod instruction ... */
690 i_instr_record *map_Mod = &records[n_records++].i_instr;
692 tp = new_type_method(ID("rt_iMod"), 2, 1);
693 set_method_param_type(tp, 0, int_tp);
694 set_method_param_type(tp, 1, int_tp);
695 set_method_res_type(tp, 0, int_tp);
697 rt_iMod.ent = new_entity(get_glob_type(), ID("__modsi3"), tp);
698 set_entity_ld_ident(rt_iMod.ent, ID("__modsi3"));
699 rt_iMod.mode = mode_T;
700 rt_iMod.res_mode = mode_Is;
701 rt_iMod.mem_proj_nr = pn_Mod_M;
702 rt_iMod.regular_proj_nr = pn_Mod_X_regular;
703 rt_iMod.exc_proj_nr = pn_Mod_X_except;
704 rt_iMod.exc_mem_proj_nr = pn_Mod_M;
705 rt_iMod.res_proj_nr = pn_Mod_res;
707 set_entity_visibility(rt_iMod.ent, visibility_external_allocated);
709 map_Mod->kind = INTRINSIC_INSTR;
710 map_Mod->op = op_Mod;
711 map_Mod->i_mapper = (i_mapper_func)i_mapper_RuntimeCall;
712 map_Mod->ctx = &rt_iMod;
714 /* ... nor an unsigned mod. */
716 i_instr_record *map_Mod = &records[n_records++].i_instr;
718 tp = new_type_method(ID("rt_uMod"), 2, 1);
719 set_method_param_type(tp, 0, uint_tp);
720 set_method_param_type(tp, 1, uint_tp);
721 set_method_res_type(tp, 0, uint_tp);
723 rt_uMod.ent = new_entity(get_glob_type(), ID("__umodsi3"), tp);
724 set_entity_ld_ident(rt_uMod.ent, ID("__umodsi3"));
725 rt_uMod.mode = mode_T;
726 rt_uMod.res_mode = mode_Iu;
727 rt_uMod.mem_proj_nr = pn_Mod_M;
728 rt_uMod.regular_proj_nr = pn_Mod_X_regular;
729 rt_uMod.exc_proj_nr = pn_Mod_X_except;
730 rt_uMod.exc_mem_proj_nr = pn_Mod_M;
731 rt_uMod.res_proj_nr = pn_Mod_res;
733 set_entity_visibility(rt_uMod.ent, visibility_external_allocated);
735 map_Mod->kind = INTRINSIC_INSTR;
736 map_Mod->op = op_Mod;
737 map_Mod->i_mapper = (i_mapper_func)i_mapper_RuntimeCall;
738 map_Mod->ctx = &rt_uMod;
742 lower_intrinsics(records, n_records, /*part_block_used=*/0);
745 /*****************************************************************
746 * ____ _ _ _____ _____
747 * | _ \ | | | | |_ _|/ ____| /\
748 * | |_) | __ _ ___| | _____ _ __ __| | | | | (___ / \
749 * | _ < / _` |/ __| |/ / _ \ '_ \ / _` | | | \___ \ / /\ \
750 * | |_) | (_| | (__| < __/ | | | (_| | _| |_ ____) / ____ \
751 * |____/ \__,_|\___|_|\_\___|_| |_|\__,_| |_____|_____/_/ \_\
753 *****************************************************************/
755 static arm_isa_t arm_isa_template = {
757 &arm_isa_if, /* isa interface */
758 &arm_gp_regs[REG_SP], /* stack pointer */
759 &arm_gp_regs[REG_R11], /* base pointer */
760 -1, /* stack direction */
761 NULL, /* main environment */
763 5, /* reload costs */
765 0, /* use generic register names instead of SP, LR, PC */
766 ARM_FPU_ARCH_FPE, /* FPU architecture */
767 NULL, /* current code generator */
768 NULL_EMITTER, /* emitter environment */
772 * Initializes the backend ISA and opens the output file.
774 static void *arm_init(FILE *file_handle) {
775 static int inited = 0;
781 isa = xmalloc(sizeof(*isa));
782 memcpy(isa, &arm_isa_template, sizeof(*isa));
787 be_emit_init_env(&isa->emit, file_handle);
789 arm_create_opcodes();
790 arm_handle_intrinsics();
792 /* we mark referenced global entities, so we can only emit those which
793 * are actually referenced. (Note: you mustn't use the type visited flag
794 * elsewhere in the backend)
796 inc_master_type_visited();
805 * Closes the output file and frees the ISA structure.
807 static void arm_done(void *self) {
808 arm_isa_t *isa = self;
810 be_gas_emit_decls(&isa->emit, isa->arch_isa.main_env, 1);
812 be_emit_destroy_env(&isa->emit);
818 * Report the number of register classes.
819 * If we don't have fp instructions, report only GP
820 * here to speed up register allocation (and makes dumps
821 * smaller and more readable).
823 static int arm_get_n_reg_class(const void *self) {
824 const arm_isa_t *isa = self;
826 /* ARGH! is called BEFORE transform */
828 return isa->cg->have_fp_insn ? 2 : 1;
832 * Return the register class with requested index.
834 static const arch_register_class_t *arm_get_reg_class(const void *self, int i) {
836 return &arm_reg_classes[i];
840 * Get the register class which shall be used to store a value of a given mode.
841 * @param self The this pointer.
842 * @param mode The mode in question.
843 * @return A register class which can hold values of the given mode.
845 const arch_register_class_t *arm_get_reg_class_for_mode(const void *self, const ir_mode *mode) {
847 if (mode_is_float(mode))
848 return &arm_reg_classes[CLASS_arm_fpa];
850 return &arm_reg_classes[CLASS_arm_gp];
854 * Produces the type which sits between the stack args and the locals on the stack.
855 * it will contain the return address and space to store the old base pointer.
856 * @return The Firm type modeling the ABI between type.
858 static ir_type *arm_get_between_type(void *self) {
859 static ir_type *between_type = NULL;
860 static ir_entity *old_bp_ent = NULL;
863 if (between_type == NULL) {
864 ir_entity *ret_addr_ent;
865 ir_type *ret_addr_type = new_type_primitive(new_id_from_str("return_addr"), mode_P);
866 ir_type *old_bp_type = new_type_primitive(new_id_from_str("bp"), mode_P);
868 between_type = new_type_class(new_id_from_str("arm_between_type"));
869 old_bp_ent = new_entity(between_type, new_id_from_str("old_bp"), old_bp_type);
870 ret_addr_ent = new_entity(between_type, new_id_from_str("old_bp"), ret_addr_type);
872 set_entity_offset(old_bp_ent, 0);
873 set_entity_offset(ret_addr_ent, get_type_size_bytes(old_bp_type));
874 set_type_size_bytes(between_type, get_type_size_bytes(old_bp_type) + get_type_size_bytes(ret_addr_type));
882 be_abi_call_flags_bits_t flags;
883 const arch_env_t *arch_env;
884 const arch_isa_t *isa;
888 static void *arm_abi_init(const be_abi_call_t *call, const arch_env_t *arch_env, ir_graph *irg)
890 arm_abi_env_t *env = xmalloc(sizeof(env[0]));
891 be_abi_call_flags_t fl = be_abi_call_get_flags(call);
892 env->flags = fl.bits;
894 env->arch_env = arch_env;
895 env->isa = arch_env->isa;
899 static void arm_abi_dont_save_regs(void *self, pset *s)
901 arm_abi_env_t *env = self;
902 if (env->flags.try_omit_fp)
903 pset_insert_ptr(s, env->isa->bp);
909 * Build the ARM prolog
911 static const arch_register_t *arm_abi_prologue(void *self, ir_node **mem, pmap *reg_map) {
912 ir_node *keep, *store;
913 arm_abi_env_t *env = self;
914 ir_graph *irg = env->irg;
915 ir_node *block = get_irg_start_block(irg);
916 arch_register_class_t *gp = &arm_reg_classes[CLASS_arm_gp];
918 ir_node *fp = be_abi_reg_map_get(reg_map, env->isa->bp);
919 ir_node *ip = be_abi_reg_map_get(reg_map, &arm_gp_regs[REG_R12]);
920 ir_node *sp = be_abi_reg_map_get(reg_map, env->isa->sp);
921 ir_node *lr = be_abi_reg_map_get(reg_map, &arm_gp_regs[REG_LR]);
922 ir_node *pc = be_abi_reg_map_get(reg_map, &arm_gp_regs[REG_PC]);
924 if (env->flags.try_omit_fp)
927 ip = be_new_Copy(gp, irg, block, sp);
928 arch_set_irn_register(env->arch_env, ip, &arm_gp_regs[REG_R12]);
929 be_set_constr_single_reg(ip, BE_OUT_POS(0), &arm_gp_regs[REG_R12] );
931 store = new_rd_arm_StoreStackM4Inc(NULL, irg, block, sp, fp, ip, lr, pc, *mem);
933 sp = new_r_Proj(irg, block, store, env->isa->sp->reg_class->mode, pn_arm_StoreStackM4Inc_ptr);
934 arch_set_irn_register(env->arch_env, sp, env->isa->sp);
935 *mem = new_r_Proj(irg, block, store, mode_M, pn_arm_StoreStackM4Inc_M);
937 keep = be_new_CopyKeep_single(gp, irg, block, ip, sp, get_irn_mode(ip));
938 be_node_set_reg_class(keep, 1, gp);
939 arch_set_irn_register(env->arch_env, keep, &arm_gp_regs[REG_R12]);
940 be_set_constr_single_reg(keep, BE_OUT_POS(0), &arm_gp_regs[REG_R12] );
942 fp = new_rd_arm_Sub_i(NULL, irg, block, keep, get_irn_mode(fp),
943 new_tarval_from_long(4, get_irn_mode(fp)));
944 arch_set_irn_register(env->arch_env, fp, env->isa->bp);
946 be_abi_reg_map_set(reg_map, env->isa->bp, fp);
947 be_abi_reg_map_set(reg_map, &arm_gp_regs[REG_R12], keep);
948 be_abi_reg_map_set(reg_map, env->isa->sp, sp);
949 be_abi_reg_map_set(reg_map, &arm_gp_regs[REG_LR], lr);
950 be_abi_reg_map_set(reg_map, &arm_gp_regs[REG_PC], pc);
956 * Builds the ARM epilogue
958 static void arm_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_map) {
959 arm_abi_env_t *env = self;
960 ir_node *curr_sp = be_abi_reg_map_get(reg_map, env->isa->sp);
961 ir_node *curr_bp = be_abi_reg_map_get(reg_map, env->isa->bp);
962 ir_node *curr_pc = be_abi_reg_map_get(reg_map, &arm_gp_regs[REG_PC]);
963 ir_node *curr_lr = be_abi_reg_map_get(reg_map, &arm_gp_regs[REG_LR]);
965 // TODO: Activate Omit fp in epilogue
966 if (env->flags.try_omit_fp) {
967 curr_sp = be_new_IncSP(env->isa->sp, env->irg, bl, curr_sp, BE_STACK_FRAME_SIZE_SHRINK);
968 add_irn_dep(curr_sp, *mem);
970 curr_lr = be_new_CopyKeep_single(&arm_reg_classes[CLASS_arm_gp], env->irg, bl, curr_lr, curr_sp, get_irn_mode(curr_lr));
971 be_node_set_reg_class(curr_lr, 1, &arm_reg_classes[CLASS_arm_gp]);
972 arch_set_irn_register(env->arch_env, curr_lr, &arm_gp_regs[REG_LR]);
973 be_set_constr_single_reg(curr_lr, BE_OUT_POS(0), &arm_gp_regs[REG_LR] );
975 curr_pc = be_new_Copy(&arm_reg_classes[CLASS_arm_gp], env->irg, bl, curr_lr );
976 arch_set_irn_register(env->arch_env, curr_pc, &arm_gp_regs[REG_PC]);
977 be_set_constr_single_reg(curr_pc, BE_OUT_POS(0), &arm_gp_regs[REG_PC] );
978 be_node_set_flags(curr_pc, BE_OUT_POS(0), arch_irn_flags_ignore);
982 tarval *tv = new_tarval_from_long(12,mode_Iu);
983 sub12_node = new_rd_arm_Sub_i(NULL, env->irg, bl, curr_bp, mode_Iu, tv);
985 //set_arm_req_out_all(sub12_node, sub12_req);
986 arch_set_irn_register(env->arch_env, sub12_node, env->isa->sp);
987 load_node = new_rd_arm_LoadStackM3( NULL, env->irg, bl, sub12_node, *mem );
989 //set_arm_req_out(load_node, &arm_default_req_arm_gp_r11, 0);
990 //set_arm_req_out(load_node, &arm_default_req_arm_gp_sp, 1);
991 //set_arm_req_out(load_node, &arm_default_req_arm_gp_pc, 2);
992 curr_bp = new_r_Proj(env->irg, bl, load_node, env->isa->bp->reg_class->mode, pn_arm_LoadStackM3_res0);
993 curr_sp = new_r_Proj(env->irg, bl, load_node, env->isa->sp->reg_class->mode, pn_arm_LoadStackM3_res1);
994 curr_pc = new_r_Proj(env->irg, bl, load_node, mode_Iu, pn_arm_LoadStackM3_res2);
995 *mem = new_r_Proj(env->irg, bl, load_node, mode_M, pn_arm_LoadStackM3_M);
996 arch_set_irn_register(env->arch_env, curr_bp, env->isa->bp);
997 arch_set_irn_register(env->arch_env, curr_sp, env->isa->sp);
998 arch_set_irn_register(env->arch_env, curr_pc, &arm_gp_regs[REG_PC]);
1000 be_abi_reg_map_set(reg_map, env->isa->sp, curr_sp);
1001 be_abi_reg_map_set(reg_map, env->isa->bp, curr_bp);
1002 be_abi_reg_map_set(reg_map, &arm_gp_regs[REG_LR], curr_lr);
1003 be_abi_reg_map_set(reg_map, &arm_gp_regs[REG_PC], curr_pc);
1006 static const be_abi_callbacks_t arm_abi_callbacks = {
1009 arm_get_between_type,
1010 arm_abi_dont_save_regs,
1017 * Get the ABI restrictions for procedure calls.
1018 * @param self The this pointer.
1019 * @param method_type The type of the method (procedure) in question.
1020 * @param abi The abi object to be modified
1022 void arm_get_call_abi(const void *self, ir_type *method_type, be_abi_call_t *abi) {
1026 int n = get_method_n_params(method_type);
1027 be_abi_call_flags_t flags = {
1029 0, /* store from left to right */
1030 0, /* store arguments sequential */
1031 1, /* try to omit the frame pointer */
1032 1, /* the function can use any register as frame pointer */
1033 1, /* a call can take the callee's address as an immediate */
1034 0, /* IRG is a leaf function */
1035 0 /* Set to one, if there is already enough room on the stack for call args. */
1040 /* set stack parameter passing style */
1041 be_abi_call_set_flags(abi, flags, &arm_abi_callbacks);
1043 for (i = 0; i < n; i++) {
1044 /* reg = get reg for param i; */
1045 /* be_abi_call_param_reg(abi, i, reg); */
1048 be_abi_call_param_reg(abi, i, arm_get_RegParam_reg(i));
1050 be_abi_call_param_stack(abi, i, 4, 0, 0);
1053 /* set return registers */
1054 n = get_method_n_ress(method_type);
1056 assert(n <= 2 && "more than two results not supported");
1058 /* In case of 64bit returns, we will have two 32bit values */
1060 tp = get_method_res_type(method_type, 0);
1061 mode = get_type_mode(tp);
1063 assert(!mode_is_float(mode) && "two FP results not supported");
1065 tp = get_method_res_type(method_type, 1);
1066 mode = get_type_mode(tp);
1068 assert(!mode_is_float(mode) && "mixed INT, FP results not supported");
1070 be_abi_call_res_reg(abi, 0, &arm_gp_regs[REG_R0]);
1071 be_abi_call_res_reg(abi, 1, &arm_gp_regs[REG_R1]);
1072 } else if (n == 1) {
1073 const arch_register_t *reg;
1075 tp = get_method_res_type(method_type, 0);
1076 assert(is_atomic_type(tp));
1077 mode = get_type_mode(tp);
1079 reg = mode_is_float(mode) ? &arm_fpa_regs[REG_F0] : &arm_gp_regs[REG_R0];
1080 be_abi_call_res_reg(abi, 0, reg);
1084 static const void *arm_get_irn_ops(const arch_irn_handler_t *self, const ir_node *irn) {
1087 return &arm_irn_ops;
1090 const arch_irn_handler_t arm_irn_handler = {
1094 const arch_irn_handler_t *arm_get_irn_handler(const void *self) {
1096 return &arm_irn_handler;
1099 int arm_to_appear_in_schedule(void *block_env, const ir_node *irn) {
1101 if(!is_arm_irn(irn))
1108 * Initializes the code generator interface.
1110 static const arch_code_generator_if_t *arm_get_code_generator_if(void *self) {
1112 return &arm_code_gen_if;
1115 list_sched_selector_t arm_sched_selector;
1118 * Returns the reg_pressure scheduler with to_appear_in_schedule() over\loaded
1120 static const list_sched_selector_t *arm_get_list_sched_selector(const void *self, list_sched_selector_t *selector) {
1123 arm_sched_selector = reg_pressure_selector;
1124 arm_sched_selector.to_appear_in_schedule = arm_to_appear_in_schedule;
1125 return &arm_sched_selector;
1128 static const ilp_sched_selector_t *arm_get_ilp_sched_selector(const void *self) {
1134 * Returns the necessary byte alignment for storing a register of given class.
1136 static int arm_get_reg_class_alignment(const void *self, const arch_register_class_t *cls) {
1139 /* ARM is a 32 bit CPU, no need for other alignment */
1143 static const be_execution_unit_t ***arm_get_allowed_execution_units(const void *self, const ir_node *irn) {
1151 static const be_machine_t *arm_get_machine(const void *self) {
1159 * Return irp irgs in the desired order.
1161 static ir_graph **arm_get_irg_list(const void *self, ir_graph ***irg_list) {
1168 * Allows or disallows the creation of Psi nodes for the given Phi nodes.
1169 * @return 1 if allowed, 0 otherwise
1171 static int arm_is_psi_allowed(ir_node *sel, ir_node *phi_list, int i, int j) {
1172 ir_node *cmp, *cmp_a, *phi;
1176 /* currently Psi support is not implemented */
1179 /* we don't want long long Psi */
1180 #define IS_BAD_PSI_MODE(mode) (!mode_is_float(mode) && get_mode_size_bits(mode) > 32)
1182 if (get_irn_mode(sel) != mode_b)
1185 cmp = get_Proj_pred(sel);
1186 cmp_a = get_Cmp_left(cmp);
1187 mode = get_irn_mode(cmp_a);
1189 if (IS_BAD_PSI_MODE(mode))
1192 /* check the Phi nodes */
1193 for (phi = phi_list; phi; phi = get_irn_link(phi)) {
1194 ir_node *pred_i = get_irn_n(phi, i);
1195 ir_node *pred_j = get_irn_n(phi, j);
1196 ir_mode *mode_i = get_irn_mode(pred_i);
1197 ir_mode *mode_j = get_irn_mode(pred_j);
1199 if (IS_BAD_PSI_MODE(mode_i) || IS_BAD_PSI_MODE(mode_j))
1203 #undef IS_BAD_PSI_MODE
1209 * Returns the libFirm configuration parameter for this backend.
1211 static const backend_params *arm_get_libfirm_params(void) {
1212 static const ir_settings_if_conv_t ifconv = {
1213 4, /* maxdepth, doesn't matter for Psi-conversion */
1214 arm_is_psi_allowed /* allows or disallows Psi creation for given selector */
1216 static ir_settings_arch_dep_t ad = {
1218 1, /* Muls are fast enough on ARM but ... */
1219 31, /* ... one shift would be possible better */
1220 NULL, /* no evaluator function */
1221 0, /* SMUL is needed, only in Arch M */
1222 0, /* UMUL is needed, only in Arch M */
1223 32, /* SMUL & UMUL available for 32 bit */
1225 static backend_params p = {
1226 1, /* need dword lowering */
1227 0, /* don't support inline assembler yet */
1228 NULL, /* no additional opcodes */
1229 NULL, /* will be set later */
1230 NULL, /* but yet no creator function */
1231 NULL, /* context for create_intrinsic_fkt */
1232 NULL, /* will be set below */
1236 p.if_conv_info = &ifconv;
1240 /* fpu set architectures. */
1241 static const lc_opt_enum_int_items_t arm_fpu_items[] = {
1242 { "softfloat", ARM_FPU_ARCH_SOFTFLOAT },
1243 { "fpe", ARM_FPU_ARCH_FPE },
1244 { "fpa", ARM_FPU_ARCH_FPA },
1245 { "vfp1xd", ARM_FPU_ARCH_VFP_V1xD },
1246 { "vfp1", ARM_FPU_ARCH_VFP_V1 },
1247 { "vfp2", ARM_FPU_ARCH_VFP_V2 },
1251 static lc_opt_enum_int_var_t arch_fpu_var = {
1252 &arm_isa_template.fpu_arch, arm_fpu_items
1255 static const lc_opt_table_entry_t arm_options[] = {
1256 LC_OPT_ENT_ENUM_INT("fpunit", "select the floating point unit", &arch_fpu_var),
1257 LC_OPT_ENT_BOOL("gen_reg_names", "use generic register names", &arm_isa_template.gen_reg_names),
1261 const arch_isa_if_t arm_isa_if = {
1264 arm_get_n_reg_class,
1266 arm_get_reg_class_for_mode,
1268 arm_get_irn_handler,
1269 arm_get_code_generator_if,
1270 arm_get_list_sched_selector,
1271 arm_get_ilp_sched_selector,
1272 arm_get_reg_class_alignment,
1273 arm_get_libfirm_params,
1274 arm_get_allowed_execution_units,
1279 void be_init_arch_arm(void)
1281 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
1282 lc_opt_entry_t *arm_grp = lc_opt_get_grp(be_grp, "arm");
1284 lc_opt_add_table(arm_grp, arm_options);
1286 be_register_isa_if("arm", &arm_isa_if);
1288 arm_init_transform();
1291 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_arch_arm);